Powers, W. T. (1973). Science 179(4071):351–356. doi:10.1126/science.179.4071.351.
Read in NOTE-963, from the reprint in Living Control Systems: Selected Papers of William T. Powers (Control Systems Group, 1989), pp. 61–78.
Key takeaways
- Behaviour controls input. A control system “will produce whatever output is required in order to achieve a constant sensed result, even in the presence of unpredictable disturbances” (reprint p. 77). Output is determined mainly by the disturbance, so stimulus-response laws follow from environmental physics once the controlled quantity is known.
- A test for controlled quantities. Disturb a candidate quantity; if the observed change is much smaller than physics predicts, it is controlled (p. 69).
- Hierarchy by reference. A higher system “opposes disturbances not by direct activation of force outputs, but by altering the reference level” of a lower one (p. 71). Its own input is still a quantity in the world.
- Intrinsic error and reorganization. The top reference levels are inherited; reorganization is “the error-driven ‘output’ of a basic inherited control system” (p. 73), stated in one paragraph.
- No imagination. Every loop in the paper closes through the environment, down to “the effort in whole muscles, as sensed kinesthetically” (p. 73).
Why this is here
Perceptual control theory, the cybernetic account closest to U1’s wording, and a rival to CLAIM-341. Action serves perception, the reverse of active inference’s emphasis on sampling. Above the lowest level, a system’s output is a reference for another system, not action on the world, so the hierarchy is internal on its output side. But in this paper each higher system still senses a quantity in the world, and every loop bottoms out in muscle effort: it has no loop that closes without the environment. That is this paper’s position, not PCT’s: Powers’s “imagination connection”, in which a reference signal is routed back as the system’s own perception, appears in his 1960 paper and his 1988 outline, where “the control loop always (except for imagination) involves effects in the external world” (Living Control Systems, p. 279). Neither is held in the record, and the book (LIT-1179) was not reached. For CLAIM-342 and QUESTION-121, its top level of inherited intrinsic quantities, whose error drives reorganization, is a scoring channel the learned hierarchy does not set. It carries no instruction for machine-learning practice.