Contribution
A short programmatic article in Science. It argues that behaviourism’s open-loop model of stimulus and response is wrong because stimuli depend on responses as much as responses on stimuli, and replaces it with a negative feedback control unit whose sensed input, not output, is held at a reference level. It gives a test for a controlled quantity (disturb it and see whether the effect is much smaller than physics predicts), and builds a two-level hierarchy from one example (a soldier pointing at a helicopter) in which the higher system acts by setting the lower system’s reference. It sketches, in one paragraph, a top level of inherited “intrinsic reference levels” whose error drives reorganization. It is a preview of the book, which it cites as “in press” (note 2, p. 78).
Key insight
“control systems control what they sense, not really what they do” (pp. 73–74). Output varies to cancel disturbances, so “the chief determinant of output is thus the disturbance” (p. 74), and what is constant is the sensed result. Stack such units and the output of each higher unit “is not a muscle force, but a reference level (variable) for a lower-order controlled quantity” (p. 77): purposes are reference signals, and a higher purpose is pursued by changing a lower one.
Assumptions
- Negative feedback and dynamic stability. The system “must be organized for negative (not positive) feedback, and … must be dynamically stable” (p. 65). Unstable systems are set aside because normal behaviour does not show their symptoms.
- Constant sensor. “If the properties of the sensor remain constant, as we may usually assume” (p. 67), control of the sensor signal is control of the external quantity.
- No learning in the unit. “This is a continuous-variable (analog) model, without provision for learning” (p. 67). Reorganization is named only at the top of the hierarchy, and deferred: “that takes us to the verge of learning theory, which is beyond the intent of this article” (p. 73).
- A slow observer. The right time scale for observing a control system “is that on which the response to an impulse-disturbance is apparently zero” (p. 68). On it, goal-seeking is “goal-maintaining behavior”.
- Additive disturbance and output. The controlled quantity changes by g(d) + h(o), so under good control g(d) ≈ −h(o) (pp. 69–70).
- The reference’s origin is left open at first. Fig. 1’s reference is “of unspecified origin” (p. 66), answered in part by the hierarchy.
Key results
There are no data. The quantitative content is a worked example.
- The test for a controlled quantity (p. 69). Postulate q_c = f(v₁ … v_n), predict the change Δq_c = g(d) a disturbance d would cause without the system, observe the actual change Δq_c. If q_c is controlled, Δq_c/Δq_c ≪ 1. Varying the form of f to minimise the ratio finds the controlled quantity.
- The soldier’s arm (p. 70). A 1 kg sideways force would raise an uncontrolled arm about 30 degrees; the arm moves 1 degree, 1/30 of the prediction, so the soldier’s opposing force is “just a trifle under 1 kilogram”. With the arm held horizontal against a 10 kg weight, an upward push of 10 kg should drop net output to zero and 11 kg should produce “an output of 1/2 kilogram downward” (pp. 70–71; as printed).
- Two levels (pp. 71–72). Pointing at a moving helicopter keeps the angle between pointing direction and line of sight at zero, while arm position, still controlled against brief pushes, drifts with the helicopter. The second system “opposes disturbances not by direct activation of force outputs, but by altering the reference level, by means of changing the reference signal for the arm-position control system” (p. 71). It needs two lower systems, horizontal and vertical.
- Up is why, down is how (pp. 72–73). Going down: arm position, then “net muscle-generated forces”, then “the control systems in the spine which control the effort in whole muscles, as sensed kinesthetically” (p. 73). Going up, the top is “certain biologically important variables, which Ashby called critical variables and which I term intrinsic quantities” (p. 73).
- Self-determination (p. 76). “organisms are self-determined in terms of inner control of what they sense, at every level of organization except the highest level.”
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Behaviour is the control of input, not output: a control system produces whatever output keeps its sensed result at a reference | strong as a property of negative feedback; as a claim about organisms, illustrative | Fig. 1 and §Feedback Control; the soldier example |
| C2 | A controlled quantity can be identified by disturbing it and finding Δ*q_c/Δq_c ≪ 1 | moderate | method stated, p. 69; one informal example |
| C3 | Above the lowest level, a system’s output is a reference for lower systems, not a muscle force | moderate | one two-level example, pp. 71–72; extension “indefinitely” asserted, p. 72 |
| C4 | Given the controlled quantity, stimulus-response laws follow from environmental physics, not from the organism | moderate | derivation g(d) ≈ −h(o), p. 74; Fig. 1 caption |
| C5 | The highest reference levels are inherited intrinsic quantities, and reorganization is the error-driven output of an inherited control system | weak | asserted in one paragraph, p. 73, citing the 1960 paper |
| C6 | Operant conditioning of humans by humans is doubtful in the long run, since people learn to circumvent arbitrary deprivation | weak | informal argument, pp. 76–77 |
Concepts
- Act and result. “What I term an act is a behavior that is arbitrarily left unanalyzed, while a result is defined as an understandable physical consequence of an act. Act and result are relative terms” (note 3, p. 78).
- Controlled quantity. A function q_c of environmental variables that the system keeps near a reference level against disturbance (p. 69).
- Reference level. The controlled quantity’s value “when the system’s output is totally unopposed”, equivalently the value that “results in zero error-correcting output” (p. 70). Distinguished from the internal reference signal.
- Disturbance. Effects on the input quantity independent of the system’s output; it “corresponds to ‘stimulus’” (Fig. 1 caption, p. 66).
- Intrinsic quantities. Ashby’s critical variables: “the quantities affected by deprivation and subsequent reinforcements that erase, or at least diminish, the errors caused by deprivation” (p. 73).
- Reorganization. “the error-driven ‘output’ of a basic inherited control system which is ultimately responsible for the particular structure of an adult’s behavioral control system” (p. 73).
Connections
The control unit is “modeled after Wiener’s original concept”, with the system boundary moved and the sensor signal, not the reference, made “the immediate consequence of a stimulus input” (p. 67). Goals as states a behaviour tends toward are taken from Ashby’s Design for a Brain and reread as a sluggish control system (p. 67); intrinsic quantities are Ashby’s critical variables (p. 73). Act and result follow Rozeboom and Brunswik (notes 3–5). Reorganization is referred to the 1960 General Feedback Theory paper with Clark and McFarland (note 8). The book, LIT-1179, is cited as in press (note 2). Baum and Reese’s reply in Science 181 (1973), with Powers’s answer, is reprinted next in the same volume (pp. 79–88) and was not read for this note.
Bearing on the record
- CLAIM-341 (thesis). The claim’s “Rivals” bullet quotes this
paper correctly: “behavior is the control of input, not output” (p. 77)
and “The output of a higher-order system is not a muscle force, but a
reference level (variable) for a lower-order controlled quantity” (p.
77). Two qualifications.
- Most of the hierarchy is internal, but not as the thesis’s loop is. In this paper a higher system’s input is still a quantity in the world: the pointing direction “relative to the direction of the helicopter” (p. 71). It acts inwardly, through a lower reference, but what it senses and controls is outside. Its loop runs through the lower systems and then through the environment; it “controls it … by using the whole lower-order system as its means of error prevention” (p. 72). So the paper’s hierarchy is internal on the output side only. That is not the thesis’s loop, in which a judgement reshapes the observer itself.
- Every loop here does bottom out in action. The chain of hows “extends further down, to the control systems in the spine which control the effort in whole muscles, as sensed kinesthetically” (p. 73). The paper has no loop that closes without the environment, and does not mention imagination or memory. The rival as the claim states it is therefore this paper’s. It is not all of perceptual control theory: Powers had an “imagination connection” in 1960, and in 1988 he wrote that “the control loop always (except for imagination) involves effects in the external world” (Living Control Systems, p. 279). Neither is this work; the book, LIT-1179, was not read. The claim should not cite this paper for PCT’s position on internal loops.
- The reference is changed, not the sensor. In the paper the higher level changes what the lower level aims at, never how it perceives: “If the properties of the sensor remain constant, as we may usually assume” (p. 67). The thesis’s diachronic form, judgement reshaping the observer, has no counterpart here except reorganization, which changes “the particular structure of an adult’s behavioral control system” (p. 73), sensors presumably included, though the paper does not say so.
- CLAIM-342 and QUESTION-121 (the anchor). The paper’s top level is exactly a scoring channel that the learned hierarchy does not set: inherited intrinsic reference levels, compared against intrinsic quantities “affected by deprivation”, whose error drives reorganization (p. 73). Organisms are self-determined “at every level of organization except the highest level” (p. 76). That is the structure CLAIM-342 attributes to Ashby (THEORY-324), stated by Powers in Ashby’s terms, and it is QUESTION-121’s candidate 2, the body. Two cautions. In the paper the intrinsic quantities are restored by “subsequent reinforcements”, events in the world, so the channel is bodily but its correction comes by action; and “For a human being, the ‘intrinsic reference levels’ probably specify far more than mere food or water intake. We cannot arbitrarily rule out any goal at this level-not even goals such as ‘self-actualization’” (p. 73), which widens the top level until it is no longer plainly independent of the interpretation.
- CLAIM-337 (counter). The paper supports its formal point in control terms: a higher system’s means is a lower system’s reference, so the lower system is the higher one’s proximal environment. It does not support relabelling a judgement as action: nothing in it is mental action.
- CLAIM-339 (counter). The paper’s “Only overwhelming force or insuperable obstacles can cause an organism to give up control of what it senses” (p. 76) is about action, not perception without action, and says nothing about paralysis.
- Checks of LIT-1180. Its summary is correct: “produces whatever output keeps a sensed result constant despite disturbances” matches p. 77, and the hierarchy sentence matches p. 77. Its “Why this is here” says “Unlike the thesis, PCT keeps action at the bottom of every loop”: true of this paper, which has no other kind of loop, but stated of PCT, for which it is not true without the imagination exception.
- For the anthology. No instruction for machine-learning practice.
Limitations
- A programmatic article without data; the soldier and the rat are thought experiments or informal observations.
- Learning is excluded from the model it presents, and reorganization is one paragraph.
- The hierarchy is shown for two levels; “can now be continued indefinitely” (p. 72) is asserted.
- The reading is of a reset reprint with silent corrections, read from an OCR extraction; the Science pagination was not checked.
Open questions
- Whether PCT’s imagination connection, a reference signal routed back as the system’s own perception, is the thesis’s judgement–observation loop or only its synchronic, self-satisfying limit (“every goal … appeared to be immediately satisfied”, Living Control Systems p. 20). That needs the book, LIT-1179, or the 1960 paper, which the record does not hold.
- Whether reorganization can change perceptual input functions, which would make it the PCT counterpart of the thesis’s reshaped observer.