Question About Feedback Functions

Phenomena Collective phenomena
61

Thanks Bruce, thanks Warren — both of these are useful, and they pull in slightly different directions, so let me take them one at a time.

Bruce — I think we’re standing in the same place, and you’ve put the load-bearing part in the language of control better than I threw it out. The whole thing rests on channel independence, and I’d been treating that as a property of the system: are the two records really separate, or do they quietly share an owner, a source, a failure mode? You state it as a live condition between controllers — each carrier may have to control a perception that the other is controlling well. Same thing from two sides: independence isn’t something the system has, it’s something it has to keep. And your framing names exactly where these arrangements fail. The second record almost always exists. What’s missing is a carrier holding it who isn’t also holding the first — which is the same requirement I’d been reaching for from the other end, that at least one channel has to reach the world from outside the system’s own categories. An independent carrier is what gives you that outside path. Your point and mine are one criterion, described twice.

Warren — the point about multiple channels being implicit in the hierarchy is fair, and I want to be precise about where I agree and where I’d draw a line. Redundancy inside one hierarchy buys you internal consistency: the levels agree with each other. What it can’t buy you is correspondence. A hierarchy fed corrupted perceptions will sit there reporting perfect control while the world burns — and every level agrees, because they’re all reading the same corrupted signal. So the second channel that does real work can’t be another input to the same controller. It has to reach the world by a different path and come back able to disagree. Not more channels down one hierarchy — one that closes outside it.

This isn’t only a diagram. I’ve traced the same structure across about thirty failures, and it holds its shape whether the medium is clay, a corporate ledger, or a set of model weights.

In an RLHF language model the loop closes through what the rater approves, not through the world — and at inference there’s no second channel at all, so the system is open-loop with respect to reality at the moment it speaks. It reports confidence, not correspondence. (Redefining the objective doesn’t fix this — a scalar reward or an active-inference preference prior are both internal; neither adds a measurement from outside. Same open loop.) In Greece’s pre-crisis accounts, a currency swap moved debt off the declared deficit — the swap cash-flows were a second channel the whole time, and nothing independent was carrying them against the number they should have contradicted. And four thousand years earlier, a clay tablet was kept against an independently held counter-roll for exactly one reason: so the first record couldn’t be quietly altered. The medium keeps changing. The missing piece is always the same one.

Which is really why I put a second loop on the table instead of more environment-modelling. Rick’s single loop is right about what it’s right about — it’s what does the controlling. Bruce’s paths are what it reaches through. Your physics is what makes a simulation run. None of that is wrong. But the thing that lets a control system catch its own blind spot isn’t a better internal model of the environment — it’s a second loop, independently carried, controlling the agreement between two records rather than either quantity directly. When they diverge, that disagreement is the error signal.

I’d rather be honest about the status of this than oversell it. I think two loops — properly constructed, at least one reaching outside the system — have a right to exist as an architecture. I don’t think it’s settled, and I could be wrong about the number or the construction; it needs testing, and the tests are simple ones. But the fact that the same shape turns up across four millennia and three completely different media is usually a sign the structure is real.

I’ve written the whole thing up — the case set, the AI work, and the ancient material — in one place, for anyone who wants the sources rather than my summary of them: https://perceptualcontroltheory.org/

Curious where you’d each take it from here.

Luk

62

Hi Luk, this is all helpful, and I get what you are saying in the ideal case where you can’t see the problems that might come in between the tablet and the human - how can a loop work from the agency of a clay tablet? Surely, each person has to ‘sign up’ to allow the clay tablet agency within each human’s own perceptual control hierarchy, otherwise it is just a clay tablet? What is written on a clay tablets has no agency over animals, for example. I am wondering the degree to which what you are describing corresponds with the ‘propositional system’ - the collectively agreed abstract codes within humans who allow the correspondence between symbolic code on the tablet and symbolic code representing the outputs of the changing system to have a ‘virtual agency’ within their own perceptual hierarchy (from the program level). See the sections on my consciousness articles on this, which I think I had sent you…
Talk to you soon,
Warren

63

Someone replied to your post.

| wmansell
August 25 |

  • | - |

I can’t see how the physics of optics would cover everything in the environment when a baseball fielder catches a ball. Surly the physical location of the fielder, the ball, and the forces affecting these are part of the environment necessary to model in order to complete the simulation for an observer of the simulations to confirm it works.

Of course, but we were only talking about the feedback function.

I accept that the PCT model only requires the feedback function and disturbances, but the viewer of the simulation requires more. You didn’t produce your excellent displays on mindreadings.com using only the physics of optics?

Of course not. You also have to know the trajectory of the ball. That’s produced by the laws of physics in my demo, but in my model testing I had actual data on the actual trajectory of the balls being caught.

But you’re getting kind of off-topic. This thread is about feedback functions, and you seem to think that representing a feedback function as a mathematical function was an oversimplification that didn’t take into account all the links between output and effect on controlled variable. I posted my equations about the feedback function involved in controlling the speed of a car to show that those feedback functions take into account all the relevant links, such as whether it is stick or automatic, in the feedback path from output to input. I think it would be nice if you would criticize that post directly rather than changing the subject and criticizing me for not mentioning physical things that you think should be included in my baseball catching simulations.

64

Actually, only one of us was right and you didn’t correctly describe why he was right. Hint: he was not right by showing that the basic control loop works. He was right because he showed that the complaints about the feedback function made by the other two are not right.

I never said such a thing. The number of controllers needed to accomplish a task depends on the nature of the task and the situation. In my ball catching simulation, there are two controllers because there are 2 degrees of freedom of movement, and thus two variables to be controlled. In one of my papers I described a two level hierarchy of controllers, with two control systems at each level, to account for the data I obtained in a complex control task.

And the stability of the feedback function has very little to do with the success of a control system. This is mainly because most feedback functions are very stable anyway, but if they do vary, as they do in one of the operant experiments that Bill Powers modeled, the control system takes care of that just like it takes care of any other disturbance to a controlled variable.

65

Hi Luk

You say:

In the PCT model, the loop is closed through the rater because the rater is in the world. Since RLHF is probably controlling for the rater’s answer to a question (which is a disturbance in PCT), the rater is actually the (main) part of the feedback connection between the system’s output and the controlled variable (the rater’s answer).

Best, Rick

66

No, this thread is about the distinction between feedback functions and feedback paths. You referred to the latter under the term ATENFELs.

Evidently the thread is a threat, as you said, but it shouldn’t be. Collecting examples of perceptions that people control is a high priority. Perceptions of aspects of the environment which are employed as ATENFELs (segments of feedback paths) are often among the perceptions that people control.

67

The clay tablet doesn’t have agency of course. The controller employing the clay tablet as a means of controlling loss from a shipment has agency. The higher-level comparator looking at the ship manifest and the clay tablet, with a reference value of zero difference, is in that controller. At another time and place a controller (the same or different) controls correspondence of the clay tablet with a counting-up of the goods placed on the referenced ship (the ‘bill of lading’), and at the destination port or ports the captain or another controller on the ship or ashore makes a count of delivered (unladen) goods. The latter count, or total of counts, is returned to the port of origin, where the comparison is done by a controller, not necessarily the same one who saw the goods off. These were surely collectively controlled variables and processes, with many individuals controlling them with pretty high gain.

68

This is why science is necessarily a collective enterprise.

In the contrarotulus configuration, one loop is the privileged standard and the other is either correct (the comparator reports no error) or incorrect.

In a science as a social institution, the ‘consensus’ or ‘mainstream’ statement of the facts of the science is regarded as the standard, and for divergent findings it may be difficult those holding them to find an audience or get published.

In science as a practice, the priviliged loop for a given assertion is that loop which is closed through the environment (in a way recognized within the science). If it is a fact in isolation, it’s an anomaly; if it’s framed within a statement of the science (or of a subfield in the science) which accounts also for the same facts as the consensus framing, then it’s supportive evidence.

But scientists are social animals like all mammals and many other orders. (Raihani, Nichola. (2021). The Social Instinct: How Cooperation Shaped the World. New York: Macmillan.)

[Minor note: Warren wasn’t saying the model of the environment is internal to the control system. He was saying that it is a necessary part of a realistic model of behavior in a natural environment. The properties of the very constrained artificial environment of a tracking task are trivial to represent in a model. The properties of a real baseball outfield (‘the pitch’ in Warren’s post) are more complex, but still relatively simple. A real game with multiple players, lights, strategy decisions, and so on, presents a different order of relevant environmental complexity.]

69

Yes, and I showed how feedback paths are handled in the derivation of feedback functions.

Not a threat. A disappointment. But since Bill passed away I, like Fessig, the giant in Princess Bride, am getting used to disappointment.

70

Bruce — this is the version I’d want on the record. The comparator is a controller with a reference of zero difference between two records; the tablet, the manifest, the count at the destination port are collectively controlled variables. That’s cleaner than how I put it, and it settles the agency question without needing anything exotic — the clay never controls; the controller using it does.

Your point about the classic contrarotulus is the one I want to sit with, because I think it’s where the interesting disagreement is. You’re right that historically one roll is the privileged standard and the other is checked against it. What I’m proposing is deliberately the other shape: no channel privileged, the controlled variable being the residual between them, alarm on divergence with neither declared true. My reason is exactly the failure mode I keep seeing — the moment one channel is the standard, whoever controls the standard can capture the whole check, and the comparator faithfully confirms a lie. Removing the privileged roll removes the single point of capture. It costs you something: you no longer know which channel is wrong, only that they disagree. But knowing they disagree is enough to stop and look, and it can’t be quietly switched off from inside.

And thank you for the correction on Warren — you’re right, I’d read him as putting the environment-model inside the control system, and that isn’t what he said. Noted.

Warren — I think we’re describing two different layers, and both are real. You’re asking how a lifeless tablet participates in control at all — and the answer is the propositional layer you point to: humans collectively agree that these marks stand for these quantities, and only then does the tablet carry anything. No argument from me there; the clay has no agency, and I’ve never meant it does.

What I’m describing sits one floor up from that. Given that the symbols already mean something, the question I’m after is structural: what arrangement catches a false record? And that’s two independent carriers of the same quantity, with a controller comparing them and acting on the difference. The propositional agreement is what lets the tablet mean “forty jars”; the second, independently held record is what catches it when the tablet says forty and the world holds thirty. Your layer is how the code gets its meaning. Mine is how a lie in that code gets caught. They stack; they don’t compete — and Bruce’s formulation of the comparator, I think, is the bridge between them.

Rick — you’re right on the mechanics, and I’d been loose. The rater is in the world, so the loop does close through the rater; the model controls the rater’s answer, and the rater is part of the feedback path. And you’re right that multiple controllers is ordinary PCT — two degrees of freedom, two controllers; I wasn’t claiming otherwise.

Two things I’d add, because I think they’re a different case than the one you’re describing. First, the moment: what you describe is the training loop, where the rater is present and closing it. My concern is inference — the rater is gone, the weights are fixed, and at the moment the model answers there’s no loop closing at all. Second, the variable: even during training, the controlled quantity is the rater’s approval, not the state of the world the answer is about. Those can diverge without limit — an answer can win approval and be false — so the loop is closed, but around approval rather than around correspondence. Both your controllers sit in one environment, reached one way; a drift between what they report and how the world actually stands has nothing to catch it.

That’s the whole of what I’m proposing: a second loop that reaches the same quantity by an independent path, so a divergence has somewhere to show up. It isn’t a claim about language models in particular — they’re just the most recent instance. The same structural gap, and the same fix, runs back through corporate accounts and clay tablets for four thousand years. The medium changes; the missing second channel doesn’t.

71

Just to clarify, I am disappointed in me, not in any of you.

72

Thanks for all your clarity Bruce! Like you, I’m trying to talk about the parts of the environment that are not currently represented by the feedback function or the disturbance for any one control unit, but Rick tries to make the discussion about the feedback function only…

73

OK, you caught me. I give up. I can’t keep you from talking about whatever the hell you think I’m trying to keep you from talking about. So would you please talk about that? Could you please talk about the parts of the environment that are not currently represented by the feedback function or the disturbance for anyone control unit and why that’s important. I don’t even understand what you’re talking about. Is it the string Quartet I bring into the room every time I do the basic tracking task?

74

Hee hee.

No it’s not the quarter. It’s the objective, task-independent properties of everything in the scenario of tracking task performance that is not part of the feedback function or disturbance for tracking that nonetheless affects what those equations turn out to be for any specific controlled variable for a specific agent - the hardware and software, the morphology of the body and its muscles, bones and sensors; the dimensions of the room; and in a collective task how those components connect between agents controlling for different variables.
Bruce, is that how you see the attenfels?

75

Quarter = Quartet

76

Please give me an example of one or two “task-independent properties of everything in the scenario of tracking task performance that is not part of the feedback function or disturbance for tracking that nonetheless affects what those [feedback] equations turn out to be.” And explain how they do this.

77

The software code of the tracking task.
The design of the computer mouse.

78

This doesn’t answer my question. How about first explaining how software and design of the mouse are not part of the feedback function in a tracking task?

79

So Rick, you are saying that the full detailed software code of the tracking task is part of the feedback function?

80

I didn’t say anything. I asked you to explain why “tracking task software” is an example of something that is not part of the feedback function yet affects it. But I can answer your question anyway: All code that affects the relationship between user output and cursor Position is part of the feedback function for the tracking task. This is true whether the code was explicitly written to be a part of that function or simply affects it because it is a side effect of what that code is doing in another thread, for example.