Let me clarify where I do and do not agree with PJ Eby, since we’ve been involved in some heated arguments that often seem to go nowhere.
I accept that the methods described here could work, and intend to try them myself.
I accept that all of the mechanisms involved in behavior can be restated in the form of a network of feedback loops (or a computer program, etc.).
I accept that Eby is acting as a perfect Bayesian when he says “Liar!” in response to those who claim they “gave it a try” and it didn’t work. To the extent that he has a model, that is what it obligates him to believe, and Eliezer Yudkowsky has extensively argued that you should find yourself questioning the data when it conflicts with your model.
So what’s the problem, then?
I do not accept that these predictions actually follow from, or were achieved through the insights of, viewing humans as feedback control systems. The explanations here for behavioral phenomena look like commonsense reasoning that is being shoehorned into controls terminology by clever relabeling. (ETA: Why do you need the concept of a “feedback control system” to think of the idea of running through the reasons you’re afraid of something, for example?)
For that reason, I ran the standard check to see if a model is actually constraining expectations by asking pjeby what it rules out, and, more importantly, why PCT says you shouldn’t observe such a phenomenon. And I still don’t have an answer.
(This doesn’t contradict my third point of agreement, because pjeby can believe something contradicts his model, even if, it turns out, the model he claims to believe in draws no such conclusion.)
Rather, based on this article, it looks like PCT is in the position of:
“Did PCT Method X solve your problem? Well, that’s because it reset your control references to where they should be. Did it fail? Well, that’s because PCT says that other (blankly solid, blackbox) circuts were, like, fighting it.”
Why do you need the concept of a “feedback control system” to think of the idea of running through the reasons you’re afraid of something, for example?
You don’t. As it says in the report I wrote, I’ve been teaching most of these things for years.
I ran the standard check to see if a model is actually constraining expectations by asking pjeby what it rules out, and, more importantly, why PCT says you shouldn’t observe such a phenomenon. And I still don’t have an answer.
And I’m still confused by what it is you expect to see, that I haven’t already answered. AFAIK, PCT doesn’t say that no behavior is ever generated except by control systems, it just says that control systems are an excellent model for describing how living systems generate behavior, and that we can make more accurate predictions about how a living system will behave if we know what variables it’s controlling for.
Since the full PCT model is Turing complete, what is it exactly that you are asking be “ruled out”?
Personally, I’m more interested in the things PCT rules in—that is, the things it predicts that other models don’t, such as the different timescales for “giving up” and symptom substitution. I’m not aware of any other model where this falls out so cleanly as a side effect of the model.
“Did PCT Method X solve your problem? Well, that’s because it reset your control references to where they should be. Did it fail? Well, that’s because PCT says that other (blankly solid, blackbox) circuts were, like, fighting it.”
It’s no more black-box than Ainslie’s picoeconomics. In fact, it’s considerably less black box than picoeconomics, which doesn’t do much to explain the internal structure of “interests” and “appetites”. PCT, OTOH provides a nice unboxing of those concepts into likely implementations.
Why do you need the concept of a “feedback control system” to think of the idea of running through the reasons you’re afraid of something, for example?
You don’t. As it says in the report I wrote, I’ve been teaching most of these things for years.
Then why does it get you there faster? If someone had long ago proposed to you that the body operates as a network of negative feedback controllers, would that make you more quickly reach the conclusion that “I should rationally think through the reasons I’m afraid of something”, as opposed to, say, blindly reflecting on your own conscious experience?
PCT … says that control systems are an excellent model for describing how living systems generate behavior, and that we can make more accurate predictions about how a living system will behave if we know what variables it’s controlling for.
Yes, and that’s quite a monster “if”. So far, I haven’t seen you identify—in the rationalist sense—a “variable being controlled”. That requires you to be able to explain it in terms “so simple a computer could understand it”. So far, no one can do that for any high-level behavior.
For example, judging sexiness of another person. To phrase it in terms of a feedback controller, I have to identify—again, in the rationalist sense, not just a pleasant sounding label—the reference being controlled. So, that means I have to specify all of the relevant things that affect my attaction level. Then, I have to find how the sensory data is transformed into a comparable format. Only then am I able to set up a model that shows an error signal which can drive behavior.
(Aside: note that the above can be rephased as saying that you need to find the person’s “invariants” of sexiness, i.e., the features that appear the same in the “sexiness” dimension, even despite arbitrary transformations applied to the sense data, like rotation, scaling, environment changes, etc. Not surprisingly, the Hawkins HTM model you love so much also appeals to “invariants” for their explanatory power, and yet leaves the whole concept as an unhelpful black box! At least, that’s what I got from reading On Intelligence.)
But all of these tasks are just as difficult as the original problem!
Since the full PCT model is Turing complete, what is it exactly that you are asking be “ruled out”?
Now I’m running into “inferential distance frustration”, as I’d have to explain the basics of technical explanation, what it means to really explain something, etc., i.e. all those posts by Eliezer Yudkowsky, starting back from his overcomingbias.com posts.
But suffice to say, yes, PCT is Turing complete. So is the C programming language, and so is a literal tape-and-table Turing machine. So, if you accept the Church-Turing Thesis, there must be an isomorphism between some feedback control model and the human body.
And between some C program and the human body.
And between some Turing machine and the human body.
Does this mean it is helpful to model the body as a Turing machine? No, no, no, a thousand times, NO! Because a “1” on the tape is going to map to some hideously complex set of features on a human body.
In other words, the (literal) Turing machine model of human behavior fails to simplify the process of predicting human behavior: it will explain the same things we knew before, but require a lot more complexity to do so, just like using a geocentric eypicycle model.
In OB/LW jargon, it lengthens the message needed to describe the observed data.
I claim the same thing is true of PCT: it will do little more than restate already known things, but allow it to be rephrased, with more difficulty, using controls terminology.
Personally, I’m more interested in the things PCT rules in—that is, the things it predicts that other models don’t, … I’m not aware of any other model where this falls out so cleanly as a side effect of the model.
Okay, good. Those are things that rationalists should look for. But people have a tendency to claim their theory predicted something after-the-fact, when an objective reading of it would say the the theory predicted no such thing. So I need something that can help me distinguish between:
a) “PCT predicts X, while other models do not.”
vs.
b) “PJ Eby’s positive affect toward PCT causes him to believe it implies we should observe X, while other models do not.”
A great way to settle the matter would be an objective specification of how exactly PCT generates predictions. But so far, it seems that to learn what PCT predicts, you have to pass the data up through the filter of someone who already likes PCT, and thus can freely claim the model says what they want it to say, with no one able to objectively demonstrate, “No, PCT says that shouldn’t happen.”
If someone had long ago proposed to you that the body operates as a network of negative feedback controllers, would that make you more quickly reach the conclusion that “I should rationally think through the reasons I’m afraid of something”, as opposed to, say, blindly reflecting on your own conscious experience?
Of course not; the paths between Theory and Practice are not symmetrical. In the context of my work, the usefulness of a theory like this is it provides me with a metaphorical framework to connect practical knowledge to. Instead of teaching all the dozens of principles, ideas, aphorisms, etc. that I have as isolated units, being able to link each one ot a central metaphor of controllers, levels, etc. makes communication and motivation easier.
To be perfectly fair, PCT would be useful for this purpose even if it were not a true or semi-true theory. However, all else being equal, I’d rather have something true that fits my practical observations, and PCT fits more of my practical observations than anything else. And I believe it is, in fact, true.
I’m merely stating the above so as to make it clear that if I thought it were only a metaphor, I would have no problem with saying, “it’s just a metaphor that aids education and motivation in applying certain practical observations by giving them a common conceptual framework.”
For example, judging sexiness of another person. To phrase it in terms of a feedback controller, I have to identify—again, in the rationalist sense, not just a pleasant sounding label—the reference being controlled. So, that means I have to specify all of the relevant things that affect my attaction level. Then, I have to find how the sensory data is transformed into a comparable format. Only then am I able to set up a model that shows an error signal which can drive behavior.
I still don’t see your point about this. Any model has to do the same thing, doesn’t it? So how is this a flaw of PCT?
And on a practical level, I just finished a webinar where we spent time breaking down people’s references for things like “Preparedness” and “Being a good father” and showing how to establish controllable perceptions for these things that could be honored in more than the breach. (For example, if your perceptual definition of “being a good father” is based solely on the actions of your kids rather than your actions, then you are in for some pain!)
IOW, I don’t actually see a lot of problem with reference breakdowns and even reference design, at the high-level applications for which I’m using the theory. Would I have a hard time defining “length” or “color” in terms of its referents? Sure, but I don’t really care. Powers does a good job of explaining what’s currently known about such invariants, and pointing to what research still needs to be done.
Not surprisingly, the Hawkins HTM model you love so much also appeals to “invariants” for their explanatory power, and yet leaves the whole concept as an unhelpful black box! At least, that’s what I got from reading On Intelligence.)
Did you ever look at any of the Numenta HTM software demos? AFAIK, they actually have some software that can learn the idea of “airplane” from noisy, extremely low-res pictures of them flying by. That is, HTMs can learn invariants from combinations of features. I’m not sure if they have any 3D rotation stuff, but the HTM model appears to explain how it could be done.
I claim the same thing is true of PCT: it will do little more than restate already known things, but allow it to be rephrased, with more difficulty, using controls terminology.
And I’ve already pointed out why this claim is false, since the controller hierarchy/time-scale correlation has already been a big help to me in my work; it was not something that was predicted by any other model of human behavior.
But so far, it seems that to learn what PCT predicts, you have to pass the data up through the filter of someone who already likes PCT, and thus can freely claim the model says what they want it to say, with no one able to objectively demonstrate, “No, PCT says that shouldn’t happen.”
Or, you could just go RTFM, instead of asking people to summarize 300 pages in a comment for you… Or you could just wait until someone you trust gives you a summary. But if you don’t trust anyone who holds a positive attitude about PCT, why do you insist on asking more questions? As I said, if you want all the detailed evidence and models, you’re eventually going to be asking me for virtually every chapter in B:CP.
What I’m teaching to my group is only a watered-down version of the highest levels, specifically as a framework to clarify, connect, and enhance things I’ve already been teaching. So my writings on it are really not the place to be looking for the math and the science.
Not surprisingly, the Hawkins HTM model you love so much also appeals to “invariants” for their explanatory power, and yet leaves the whole concept as an unhelpful black box! At least, that’s what I got from reading On Intelligence.)
Did you ever look at any of the Numenta HTM software demos? AFAIK, they actually have some software that can learn the idea of “airplane” from noisy, extremely low-res pictures of them flying by.
Actually, yes, I have downloaded their demos, expecting to be wowed, but then fell over laughing. Specifically, this one. It claims to be able to learn to recognize simple black/white 16x16x pixel images using HTM and saccading the images around. But then I gave it a spin, had it learn the images, and then tested it by drawing one of the figures with a very, very slight rotation, which completely screwed up its ability to identify it.
Not impressive.
I claim the same thing is true of PCT: it will do little more than restate already known things, but allow it to be rephrased, with more difficulty, using controls terminology.
And I’ve already pointed out why this claim is false, since the controller hierarchy/time-scale correlation has already been a big help to me in my work; it was not something that was predicted by any other model of human behavior.
No, what you have shown is that you learned of PCT and HTM, and then you believe you improved in your work. As per my two previous comments in this thread, I can (edited phrasing) accept both of those claims and still doubt the more interesting claims, specifically, that the model actually did help, rather than you merely thinking it did because you could rephrase your intuitive, commonsense reasoning in the model’s terminology. I could also doubt that your ability to help people improved.
Or, you could just go RTFM, instead of asking people to summarize 300 pages in a comment for you… As I said, if you want all the detailed evidence and models, you’re eventually going to be asking me for virtually every chapter in B:CP.
Please pay attention. I am R-ingTFM, and I even complained that one of the Powers demos understated the strength of their point about feedback control. I already told you I’m going to try your advice. I’ve read several of the pdfs you’ve linked, including TheSelfHelpMyth.pdf linked here, and will read several more, and probably even buy Behavior. (Though I couldn’t get the freebie you mentioned to work because the website crapped out after I entered my info). I am making every effort to consider this model.
But it is simply not acceptable of you to act like the only alternatives are to repeat hundreds of pages, or speak in dumbed-down blackbox terminology. You can e.g. summarize the chain of useful, critical insights that get me from “it’s a network of feedback controllers” to a useful model, so I know which part I’d be skeptical of and which parts assume the solution of problems I know to be unsolved, so I know where to direct my attention.
drawing one of the figures with a very, very slight rotation, which completely screwed up its ability to identify it.
I’m not clear on whether you took this bit from their docs into account:
The system was NOT trained on upside down images, or rotations and skews
beyond a simple right-to-left flip. In addition, the system was not trained on any curved lines, only straight line objects.
That is, I’m not clear whether the steps you’re describing include training on rotations or not.
rather than you merely thinking it did because you could rephrase your intuitive, commonsense reasoning in the model’s terminology
No, I gave you one specific prediction that PCT makes: higher-level controllers operate over longer time scales than low-level ones. This prediction is not a part of any other model I know of. Do you know of another model that makes this prediction? I only know of models that basically say that symptom substitution takes time, with no explanation of how it occurs.
This doesn’t have anything to do with whether I believe that prediction to be useful; the prediction is still there, the observation that people do it is still there, and the lack of explanation of that fact is still there, even if you remove me from the picture entirely.
You can e.g. summarize the chain of useful, critical insights that get me from “it’s a network of feedback controllers” to a useful model, so I know which part I’d be skeptical of and which parts assume the solution of problems I know to be unsolved, so I know where to direct my attention.
I can only do that if I understand specifically what it is you don’t get—and I still don’t.
For example, I don’t see why the existence of unsolved problems is a problem, or even remotely relevant, if all the other models we have have to make the same assumption.
From my POV, you are ignoring the things that make PCT useful: namely that it actually predicts as normal, things that other current behavioral models have to treat as special cases or try to handwave out of existence. It’s not that PCT is “simpler” than stimulus-response or “action steps” models, it’s that it’s the simplest model that improves on our ability to make correct predictions about behavior.
Your argument seems to be, “but PCT requires us to gather more information in order to make those predictions”. And my answer to that is, “So what? Once you have that information, you can make way better predictions.” And it’s not that you could just feed the same information into some other model and get similar predictions—the other models don’t even tell you what experiments to perform to get yes-or-no answers.
To put it another way, to the extent that PCT requires you to be more precise or gather more information, it is doing so because that degree of actual uncertainty or lack of knowledge exists… and current experimental models disguise that lack of understanding behind statistics.
In contrast, to do a PCT experiment, you need to have a more-specific, falsifiable hypothesis: is the animal or person controlling quantity X or not? You may have to do more experiments in order to identify the correct “X”, but you will actually know something real, rather than, “47% of rats appear to do Y in the presence of Z”.
That is, I’m not clear whether the steps you’re describing include training on rotations or not.
But that’s a pretty basic transformation, and if they could handle it, they would have done so. In any case, the rotation was very slight, and was only one of many tests I gave it. It didn’t merely assign a slightly lower probability to the correct answer, it fell off the list entirely.
Consider how tiny the pictures are, this is not encouraging.
Your argument seems to be, “but PCT requires us to gather more information in order to make those predictions”. And my answer to that is, “So what? Once you have that information, you can make way better predictions.”
No, you misunderstand: my complaint is that PCT requires us to solve problems of equal or greater difficulty than the initial problem being solved. To better explain what I mean, I gave you the example with the literal tape-and-table Turing machine. Watch what happens when I make your same point, but in advocacy of the “Turing machine model of human behavior”.
“I’ve discovered a great insight that helps unify my research and better assist people with their problems. It’s to view them as a long, sectioned tape with a reader and state record, which [explanation of Turing machine]. This model is so useful because all I have to do is find out whether people have 1′s rather than 0′s in places 5000-5500 on their tape, and if they do, I just have to change state 4000 to erase rather than merely move state! This helps explain why people have trouble in their lives, because they don’t erase bad memories.”
See the problems with my version?
1) Any model of a human as a Turing machine would be way more complex than the phenomenon I’m trying to explain, so the insight it gives is imaginary.
2) Even given a working model, the mapping from any part of the TM model to the human is hideously complex.
3) “Finding someone’s 1′s and 0′s” is near impossible because of the complexity of the mapping.
4) The analogy between erasing memories and erasure operations is only superficial, and not indicative of the model’s strength.
5) Because I obviously could not have a TM model of humans, I’m not actually getting my insight from the model, but from somewhere else.
And points 1-5 are exactly what I claim is going on with you and PCT.
Nevertheless, I will confess I’ve gotten more interested in PCT, and it definitely looks scientific for the low level systems. I’ve read the first two Byte magazine articles and reproduced it in Matlab’s Simulink, and I’m now reading the third, where it introduces hierarchies.
My main dispute is with your insistence that you can already usefully apply real predictions from PCT at higher-level systems, where the parallels with feedback control systems appear very superficial and the conclusions seem to be reached with commonsense reasoning unaided by PCT.
Btw: my apologies, but somehow I accidentally deleted a part of my last reply before posting it, and my remark now resides only in my memory. It’s related to the same point I just made. I’ll put it here so you don’t need to reply a second time to that post:
To phrase it in terms of a feedback controller, I have to identify—again, in the rationalist sense, not just a pleasant sounding label—the reference being controlled. So, that means I have to specify all of the relevant things that affect my attaction level. Then, I have to find how the sensory data is transformed into a comparable format … Only then am I able to set up a model that shows an error signal which can drive behavior.
I still don’t see your point about this. Any model has to do the same thing, doesn’t it? So how is this a flaw of PCT?
No, a model doesn’t need to do the same thing. A purely neuronal model would not need to have the concept of “sexiness” and a comparator for it. Remember, the whole framing of a situation as a “romantic relationship” is just that: a framing the we have imposed on it to make sense of the world. It does not exist at lower levels, and so models need not be able to indentify such complex “invariants”.
I’m sorry, but I’m still utterly baffled by your comments, since your proposed “purely neuronal” model is more analogous to the Turing machine.
It sounds a bit like the part you’re missing is the PCT experimental design philosophy, aka the Test—a way of formulating and testing control hypotheses at arbitrary levels of the hierarchy. To test “sexiness” or some other high-level value, it is not necessary to completely specify all its lower-level components, unless of course the goal of your experiment is to identify those components.
We don’t need, for example, to break down how object invariance happens to be able to do an experiment where a rat presses a bar! We assume the rat can identify the bar and determine whether it is currently pressed. The interesting part is what other things (like food, mate availability, shock-avoidance, whatever) that you can get the rat to control by pressing a bar. (At least, at higher levels.)
I’m sorry, but I’m still utterly baffled by your comments, since your proposed “purely neuronal” model is more analogous to the Turing machine.
So? I agree that the “purely neuronal” model would be really complex (though not as complex as the Turing machine would be). I just brought it up in order to show how a model doesn’t “need to have a sexiness comparator anyway”, so you do have to justify the simplicity gained when you posit that there is one.
It sounds a bit like the part you’re missing is the PCT experimental design philosophy, aka the Test—a way of formulating and testing control hypotheses at arbitrary levels of the hierarchy. To test “sexiness” or some other high-level value, it is not necessary to completely specify all its lower-level components, unless of course the goal of your experiment is to identify those components.
But if you don’t specify all of the lower level components, then your controls explanation is just a restating of the problem, not a simplifying of it. The insight you claim you are getting from it is actually from your commonsense reasoning. Indeed, virtually every insight you “explain” by PCT, you got some other way.
We don’t need, for example, to break down how object invariance happens to be able to do an experiment where a rat presses a bar!
Sure, but that’s because you don’t need to account for the rat’s ability to identify the bar in a wide variety of contexts and transformations, which is the entire point of looking for invariants.
But if you don’t specify all of the lower level components, then your controls explanation is just a restating of the problem, not a simplifying of it. The insight you claim you are getting from it is actually from your commonsense reasoning.
Kindly explain what “commonsense reasoning” explains the “symptom substitution” phenomenon in hypnosis, and in particular, explains why the duration of effect varies, using any model but PCT.
While I can look up “symptom substitution”, I’ll to know more specifically what you mean by this. But I’d have to be convinced that PCT explains it first in a way that doesn’t smuggle in your commonsense reasoning.
Now, if you want examples of how commonsense reasoning leads to the same conclusions that are provided as examples of the success of PCT, that I already have by the boatload. This whole top-level post is an example of using commonsense reasoning but attributing it to PCT. For example, long before I was aware of the concept of a control system, or even feedback (as such) I handled my fears (as does virtually everyone else) by thinking through what exactly it is about the feared thing that worries me.
Furthermore, it is obvious to most people that if you believe obstacles X, Y, and Z are keeping you from pursuing goal G, you should think up ways to overcome X, Y, and Z, and yet Kaj here presents that as something derived from PCT.
While I can look up “symptom substitution”, I’ll to know more specifically what you mean by this.
Specifically, find a “commonsense” explanation that explains why symptom substitution takes time to occur, without reference to PCT’s notion of a perception averaged over time.
Googling “symptom substitution” lead me to a journal article that argued that people have tried and failed to find evidence that it happens...
That’s Freudian symptom substitution, and in any case, the article is splitting hairs: it says that if you stop a child sucking its thumb, and it finds some other way to get its needs met, then that doesn’t count as “symptom substitution”. (IOW, the authors of the paper more or less defined it into nonexistence, such that if it exists and makes sense, it’s not symptom substitution!)
Also, the paper raises the same objection to the Freudian model of symptom substitution that I do: namely, that there is no explanation of the time frame factor.
In contrast, PCT unifies the cases both ruled-in and ruled out by this paper, and offers a better explanation for the varying time frame issue, in that the time frame is governed by the perceptual decay of the controlled variable.
I purchased Behavior: The Control of Perception and am reading it. Unless someone else does so first, I plan to write a review of it for LW. A key point is that at least part of PCT is actually right. The lowest level controllers, such as those controlling tendon tension, are verifiably there. So far as I can see so far, real physical structures corresponding pretty closely to second and third level controllers also exist and have been pointed to by science. I haven’t gotten further than this yet, but teasers within the book indicate that (when the book was written of course) there is good evidence that some fifth level control systems exist in particular places in the brain, and thus fourth level somewhere. Whether it’s control systems (or something closish to them) all the way up? Dunno. But the first couple levels, not yet into the realm of traditional psychology or whatnot, those definitely exist in humans. And animals of course. The description of the scattershot electrodes in hundreds of cats experiment was pretty interesting. :)
That said, you’re absolutely right, there should be some definite empirical implications of the theory. For example due to the varying length of the paths at various supposed levels, it should be possible to devise an experiment around a simple tracking task with carefully selected disturbances which will have one predicted result under PCT and another under some other model. Also, predicting specific discretization of tasks that look continuous should be possible… I have not spent a lot of time thinking about how to devise a test like this yet, unfortunately.
Please add PCT to the wiki as Jargon and link there when this term, whatever it means, is used for the first time in a thread. It is not in the first 10 Google hits.
It seems jimrandomh has taken the time to do so; the wikipedia article should be helpful.
In defense of people using the acronym without definition, though, it seemed fairly obvious if you look at the wikipedia disambig page for the acronym in question.
Whether it’s control systems (or something closish to them) all the way up? Dunno. But the first couple levels, not yet into the realm of traditional psychology or whatnot, those definitely exist in humans.
As a general-purpose prior assumption for systems designed by evolutionary processes, reusing or adapting existing systems is far more likely than spontaneous creation of new systems.
Thus, if it can be demonstrated that a model accurately represents low-level hierarchical systems, this is reasonably good evidence in favor of that model applying all the way to the top levels as opposed to other models with similar explanatory power for said upper levels.
Let me clarify where I do and do not agree with PJ Eby, since we’ve been involved in some heated arguments that often seem to go nowhere.
I accept that the methods described here could work, and intend to try them myself.
I accept that all of the mechanisms involved in behavior can be restated in the form of a network of feedback loops (or a computer program, etc.).
I accept that Eby is acting as a perfect Bayesian when he says “Liar!” in response to those who claim they “gave it a try” and it didn’t work. To the extent that he has a model, that is what it obligates him to believe, and Eliezer Yudkowsky has extensively argued that you should find yourself questioning the data when it conflicts with your model.
So what’s the problem, then?
I do not accept that these predictions actually follow from, or were achieved through the insights of, viewing humans as feedback control systems. The explanations here for behavioral phenomena look like commonsense reasoning that is being shoehorned into controls terminology by clever relabeling. (ETA: Why do you need the concept of a “feedback control system” to think of the idea of running through the reasons you’re afraid of something, for example?)
For that reason, I ran the standard check to see if a model is actually constraining expectations by asking pjeby what it rules out, and, more importantly, why PCT says you shouldn’t observe such a phenomenon. And I still don’t have an answer.
(This doesn’t contradict my third point of agreement, because pjeby can believe something contradicts his model, even if, it turns out, the model he claims to believe in draws no such conclusion.)
Rather, based on this article, it looks like PCT is in the position of:
“Did PCT Method X solve your problem? Well, that’s because it reset your control references to where they should be. Did it fail? Well, that’s because PCT says that other (blankly solid, blackbox) circuts were, like, fighting it.”
You don’t. As it says in the report I wrote, I’ve been teaching most of these things for years.
And I’m still confused by what it is you expect to see, that I haven’t already answered. AFAIK, PCT doesn’t say that no behavior is ever generated except by control systems, it just says that control systems are an excellent model for describing how living systems generate behavior, and that we can make more accurate predictions about how a living system will behave if we know what variables it’s controlling for.
Since the full PCT model is Turing complete, what is it exactly that you are asking be “ruled out”?
Personally, I’m more interested in the things PCT rules in—that is, the things it predicts that other models don’t, such as the different timescales for “giving up” and symptom substitution. I’m not aware of any other model where this falls out so cleanly as a side effect of the model.
It’s no more black-box than Ainslie’s picoeconomics. In fact, it’s considerably less black box than picoeconomics, which doesn’t do much to explain the internal structure of “interests” and “appetites”. PCT, OTOH provides a nice unboxing of those concepts into likely implementations.
Then why does it get you there faster? If someone had long ago proposed to you that the body operates as a network of negative feedback controllers, would that make you more quickly reach the conclusion that “I should rationally think through the reasons I’m afraid of something”, as opposed to, say, blindly reflecting on your own conscious experience?
Yes, and that’s quite a monster “if”. So far, I haven’t seen you identify—in the rationalist sense—a “variable being controlled”. That requires you to be able to explain it in terms “so simple a computer could understand it”. So far, no one can do that for any high-level behavior.
For example, judging sexiness of another person. To phrase it in terms of a feedback controller, I have to identify—again, in the rationalist sense, not just a pleasant sounding label—the reference being controlled. So, that means I have to specify all of the relevant things that affect my attaction level. Then, I have to find how the sensory data is transformed into a comparable format. Only then am I able to set up a model that shows an error signal which can drive behavior.
(Aside: note that the above can be rephased as saying that you need to find the person’s “invariants” of sexiness, i.e., the features that appear the same in the “sexiness” dimension, even despite arbitrary transformations applied to the sense data, like rotation, scaling, environment changes, etc. Not surprisingly, the Hawkins HTM model you love so much also appeals to “invariants” for their explanatory power, and yet leaves the whole concept as an unhelpful black box! At least, that’s what I got from reading On Intelligence.)
But all of these tasks are just as difficult as the original problem!
Now I’m running into “inferential distance frustration”, as I’d have to explain the basics of technical explanation, what it means to really explain something, etc., i.e. all those posts by Eliezer Yudkowsky, starting back from his overcomingbias.com posts.
But suffice to say, yes, PCT is Turing complete. So is the C programming language, and so is a literal tape-and-table Turing machine. So, if you accept the Church-Turing Thesis, there must be an isomorphism between some feedback control model and the human body.
And between some C program and the human body.
And between some Turing machine and the human body.
Does this mean it is helpful to model the body as a Turing machine? No, no, no, a thousand times, NO! Because a “1” on the tape is going to map to some hideously complex set of features on a human body.
In other words, the (literal) Turing machine model of human behavior fails to simplify the process of predicting human behavior: it will explain the same things we knew before, but require a lot more complexity to do so, just like using a geocentric eypicycle model.
In OB/LW jargon, it lengthens the message needed to describe the observed data.
I claim the same thing is true of PCT: it will do little more than restate already known things, but allow it to be rephrased, with more difficulty, using controls terminology.
Okay, good. Those are things that rationalists should look for. But people have a tendency to claim their theory predicted something after-the-fact, when an objective reading of it would say the the theory predicted no such thing. So I need something that can help me distinguish between:
a) “PCT predicts X, while other models do not.”
vs.
b) “PJ Eby’s positive affect toward PCT causes him to believe it implies we should observe X, while other models do not.”
A great way to settle the matter would be an objective specification of how exactly PCT generates predictions. But so far, it seems that to learn what PCT predicts, you have to pass the data up through the filter of someone who already likes PCT, and thus can freely claim the model says what they want it to say, with no one able to objectively demonstrate, “No, PCT says that shouldn’t happen.”
Of course not; the paths between Theory and Practice are not symmetrical. In the context of my work, the usefulness of a theory like this is it provides me with a metaphorical framework to connect practical knowledge to. Instead of teaching all the dozens of principles, ideas, aphorisms, etc. that I have as isolated units, being able to link each one ot a central metaphor of controllers, levels, etc. makes communication and motivation easier.
To be perfectly fair, PCT would be useful for this purpose even if it were not a true or semi-true theory. However, all else being equal, I’d rather have something true that fits my practical observations, and PCT fits more of my practical observations than anything else. And I believe it is, in fact, true.
I’m merely stating the above so as to make it clear that if I thought it were only a metaphor, I would have no problem with saying, “it’s just a metaphor that aids education and motivation in applying certain practical observations by giving them a common conceptual framework.”
I still don’t see your point about this. Any model has to do the same thing, doesn’t it? So how is this a flaw of PCT?
And on a practical level, I just finished a webinar where we spent time breaking down people’s references for things like “Preparedness” and “Being a good father” and showing how to establish controllable perceptions for these things that could be honored in more than the breach. (For example, if your perceptual definition of “being a good father” is based solely on the actions of your kids rather than your actions, then you are in for some pain!)
IOW, I don’t actually see a lot of problem with reference breakdowns and even reference design, at the high-level applications for which I’m using the theory. Would I have a hard time defining “length” or “color” in terms of its referents? Sure, but I don’t really care. Powers does a good job of explaining what’s currently known about such invariants, and pointing to what research still needs to be done.
Did you ever look at any of the Numenta HTM software demos? AFAIK, they actually have some software that can learn the idea of “airplane” from noisy, extremely low-res pictures of them flying by. That is, HTMs can learn invariants from combinations of features. I’m not sure if they have any 3D rotation stuff, but the HTM model appears to explain how it could be done.
And I’ve already pointed out why this claim is false, since the controller hierarchy/time-scale correlation has already been a big help to me in my work; it was not something that was predicted by any other model of human behavior.
Or, you could just go RTFM, instead of asking people to summarize 300 pages in a comment for you… Or you could just wait until someone you trust gives you a summary. But if you don’t trust anyone who holds a positive attitude about PCT, why do you insist on asking more questions? As I said, if you want all the detailed evidence and models, you’re eventually going to be asking me for virtually every chapter in B:CP.
What I’m teaching to my group is only a watered-down version of the highest levels, specifically as a framework to clarify, connect, and enhance things I’ve already been teaching. So my writings on it are really not the place to be looking for the math and the science.
Actually, yes, I have downloaded their demos, expecting to be wowed, but then fell over laughing. Specifically, this one. It claims to be able to learn to recognize simple black/white 16x16x pixel images using HTM and saccading the images around. But then I gave it a spin, had it learn the images, and then tested it by drawing one of the figures with a very, very slight rotation, which completely screwed up its ability to identify it.
Not impressive.
No, what you have shown is that you learned of PCT and HTM, and then you believe you improved in your work. As per my two previous comments in this thread, I can (edited phrasing) accept both of those claims and still doubt the more interesting claims, specifically, that the model actually did help, rather than you merely thinking it did because you could rephrase your intuitive, commonsense reasoning in the model’s terminology. I could also doubt that your ability to help people improved.
Please pay attention. I am R-ingTFM, and I even complained that one of the Powers demos understated the strength of their point about feedback control. I already told you I’m going to try your advice. I’ve read several of the pdfs you’ve linked, including TheSelfHelpMyth.pdf linked here, and will read several more, and probably even buy Behavior. (Though I couldn’t get the freebie you mentioned to work because the website crapped out after I entered my info). I am making every effort to consider this model.
But it is simply not acceptable of you to act like the only alternatives are to repeat hundreds of pages, or speak in dumbed-down blackbox terminology. You can e.g. summarize the chain of useful, critical insights that get me from “it’s a network of feedback controllers” to a useful model, so I know which part I’d be skeptical of and which parts assume the solution of problems I know to be unsolved, so I know where to direct my attention.
I’m not clear on whether you took this bit from their docs into account:
That is, I’m not clear whether the steps you’re describing include training on rotations or not.
No, I gave you one specific prediction that PCT makes: higher-level controllers operate over longer time scales than low-level ones. This prediction is not a part of any other model I know of. Do you know of another model that makes this prediction? I only know of models that basically say that symptom substitution takes time, with no explanation of how it occurs.
This doesn’t have anything to do with whether I believe that prediction to be useful; the prediction is still there, the observation that people do it is still there, and the lack of explanation of that fact is still there, even if you remove me from the picture entirely.
I can only do that if I understand specifically what it is you don’t get—and I still don’t.
For example, I don’t see why the existence of unsolved problems is a problem, or even remotely relevant, if all the other models we have have to make the same assumption.
From my POV, you are ignoring the things that make PCT useful: namely that it actually predicts as normal, things that other current behavioral models have to treat as special cases or try to handwave out of existence. It’s not that PCT is “simpler” than stimulus-response or “action steps” models, it’s that it’s the simplest model that improves on our ability to make correct predictions about behavior.
Your argument seems to be, “but PCT requires us to gather more information in order to make those predictions”. And my answer to that is, “So what? Once you have that information, you can make way better predictions.” And it’s not that you could just feed the same information into some other model and get similar predictions—the other models don’t even tell you what experiments to perform to get yes-or-no answers.
To put it another way, to the extent that PCT requires you to be more precise or gather more information, it is doing so because that degree of actual uncertainty or lack of knowledge exists… and current experimental models disguise that lack of understanding behind statistics.
In contrast, to do a PCT experiment, you need to have a more-specific, falsifiable hypothesis: is the animal or person controlling quantity X or not? You may have to do more experiments in order to identify the correct “X”, but you will actually know something real, rather than, “47% of rats appear to do Y in the presence of Z”.
But that’s a pretty basic transformation, and if they could handle it, they would have done so. In any case, the rotation was very slight, and was only one of many tests I gave it. It didn’t merely assign a slightly lower probability to the correct answer, it fell off the list entirely.
Consider how tiny the pictures are, this is not encouraging.
No, you misunderstand: my complaint is that PCT requires us to solve problems of equal or greater difficulty than the initial problem being solved. To better explain what I mean, I gave you the example with the literal tape-and-table Turing machine. Watch what happens when I make your same point, but in advocacy of the “Turing machine model of human behavior”.
“I’ve discovered a great insight that helps unify my research and better assist people with their problems. It’s to view them as a long, sectioned tape with a reader and state record, which [explanation of Turing machine]. This model is so useful because all I have to do is find out whether people have 1′s rather than 0′s in places 5000-5500 on their tape, and if they do, I just have to change state 4000 to erase rather than merely move state! This helps explain why people have trouble in their lives, because they don’t erase bad memories.”
See the problems with my version?
1) Any model of a human as a Turing machine would be way more complex than the phenomenon I’m trying to explain, so the insight it gives is imaginary.
2) Even given a working model, the mapping from any part of the TM model to the human is hideously complex.
3) “Finding someone’s 1′s and 0′s” is near impossible because of the complexity of the mapping.
4) The analogy between erasing memories and erasure operations is only superficial, and not indicative of the model’s strength.
5) Because I obviously could not have a TM model of humans, I’m not actually getting my insight from the model, but from somewhere else.
And points 1-5 are exactly what I claim is going on with you and PCT.
Nevertheless, I will confess I’ve gotten more interested in PCT, and it definitely looks scientific for the low level systems. I’ve read the first two Byte magazine articles and reproduced it in Matlab’s Simulink, and I’m now reading the third, where it introduces hierarchies.
My main dispute is with your insistence that you can already usefully apply real predictions from PCT at higher-level systems, where the parallels with feedback control systems appear very superficial and the conclusions seem to be reached with commonsense reasoning unaided by PCT.
Btw: my apologies, but somehow I accidentally deleted a part of my last reply before posting it, and my remark now resides only in my memory. It’s related to the same point I just made. I’ll put it here so you don’t need to reply a second time to that post:
No, a model doesn’t need to do the same thing. A purely neuronal model would not need to have the concept of “sexiness” and a comparator for it. Remember, the whole framing of a situation as a “romantic relationship” is just that: a framing the we have imposed on it to make sense of the world. It does not exist at lower levels, and so models need not be able to indentify such complex “invariants”.
I’m sorry, but I’m still utterly baffled by your comments, since your proposed “purely neuronal” model is more analogous to the Turing machine.
It sounds a bit like the part you’re missing is the PCT experimental design philosophy, aka the Test—a way of formulating and testing control hypotheses at arbitrary levels of the hierarchy. To test “sexiness” or some other high-level value, it is not necessary to completely specify all its lower-level components, unless of course the goal of your experiment is to identify those components.
We don’t need, for example, to break down how object invariance happens to be able to do an experiment where a rat presses a bar! We assume the rat can identify the bar and determine whether it is currently pressed. The interesting part is what other things (like food, mate availability, shock-avoidance, whatever) that you can get the rat to control by pressing a bar. (At least, at higher levels.)
So? I agree that the “purely neuronal” model would be really complex (though not as complex as the Turing machine would be). I just brought it up in order to show how a model doesn’t “need to have a sexiness comparator anyway”, so you do have to justify the simplicity gained when you posit that there is one.
But if you don’t specify all of the lower level components, then your controls explanation is just a restating of the problem, not a simplifying of it. The insight you claim you are getting from it is actually from your commonsense reasoning. Indeed, virtually every insight you “explain” by PCT, you got some other way.
Sure, but that’s because you don’t need to account for the rat’s ability to identify the bar in a wide variety of contexts and transformations, which is the entire point of looking for invariants.
Kindly explain what “commonsense reasoning” explains the “symptom substitution” phenomenon in hypnosis, and in particular, explains why the duration of effect varies, using any model but PCT.
While I can look up “symptom substitution”, I’ll to know more specifically what you mean by this. But I’d have to be convinced that PCT explains it first in a way that doesn’t smuggle in your commonsense reasoning.
Now, if you want examples of how commonsense reasoning leads to the same conclusions that are provided as examples of the success of PCT, that I already have by the boatload. This whole top-level post is an example of using commonsense reasoning but attributing it to PCT. For example, long before I was aware of the concept of a control system, or even feedback (as such) I handled my fears (as does virtually everyone else) by thinking through what exactly it is about the feared thing that worries me.
Furthermore, it is obvious to most people that if you believe obstacles X, Y, and Z are keeping you from pursuing goal G, you should think up ways to overcome X, Y, and Z, and yet Kaj here presents that as something derived from PCT.
Specifically, find a “commonsense” explanation that explains why symptom substitution takes time to occur, without reference to PCT’s notion of a perception averaged over time.
Googling “symptom substitution” lead me to a journal article that argued that people have tried and failed to find evidence that it happens...
That’s Freudian symptom substitution, and in any case, the article is splitting hairs: it says that if you stop a child sucking its thumb, and it finds some other way to get its needs met, then that doesn’t count as “symptom substitution”. (IOW, the authors of the paper more or less defined it into nonexistence, such that if it exists and makes sense, it’s not symptom substitution!)
Also, the paper raises the same objection to the Freudian model of symptom substitution that I do: namely, that there is no explanation of the time frame factor.
In contrast, PCT unifies the cases both ruled-in and ruled out by this paper, and offers a better explanation for the varying time frame issue, in that the time frame is governed by the perceptual decay of the controlled variable.
I purchased Behavior: The Control of Perception and am reading it. Unless someone else does so first, I plan to write a review of it for LW. A key point is that at least part of PCT is actually right. The lowest level controllers, such as those controlling tendon tension, are verifiably there. So far as I can see so far, real physical structures corresponding pretty closely to second and third level controllers also exist and have been pointed to by science. I haven’t gotten further than this yet, but teasers within the book indicate that (when the book was written of course) there is good evidence that some fifth level control systems exist in particular places in the brain, and thus fourth level somewhere. Whether it’s control systems (or something closish to them) all the way up? Dunno. But the first couple levels, not yet into the realm of traditional psychology or whatnot, those definitely exist in humans. And animals of course. The description of the scattershot electrodes in hundreds of cats experiment was pretty interesting. :)
That said, you’re absolutely right, there should be some definite empirical implications of the theory. For example due to the varying length of the paths at various supposed levels, it should be possible to devise an experiment around a simple tracking task with carefully selected disturbances which will have one predicted result under PCT and another under some other model. Also, predicting specific discretization of tasks that look continuous should be possible… I have not spent a lot of time thinking about how to devise a test like this yet, unfortunately.
Please add PCT to the wiki as Jargon and link there when this term, whatever it means, is used for the first time in a thread. It is not in the first 10 Google hits.
It seems jimrandomh has taken the time to do so; the wikipedia article should be helpful.
In defense of people using the acronym without definition, though, it seemed fairly obvious if you look at the wikipedia disambig page for the acronym in question.
As a general-purpose prior assumption for systems designed by evolutionary processes, reusing or adapting existing systems is far more likely than spontaneous creation of new systems.
Thus, if it can be demonstrated that a model accurately represents low-level hierarchical systems, this is reasonably good evidence in favor of that model applying all the way to the top levels as opposed to other models with similar explanatory power for said upper levels.