Yeah, is it too much to ask that in a two-line top-level post, over half of which is transcribed to get stuff like that right?
Anyway, I’d like to self-servingly nominate this recent exchange as a case study in more advanced baloney detection, where someone’s supercool, unfairly suppressed Theory of Everything turns out to be, “Um, find the solution some other way, and then shoehorn it into my theory.”
It is not supercool, it is not suppressed, it is not a theory of everything, and there is no shoehorning. It is merely a theory of living organisms that works where it has been applied, and it has been applied.
But everyone, feel free to apply Shermer’s advice, or whatever more advanced techniques of rationality you have, to the matter. I hope for no less in posting here.
When your theory’s “explanation” of a phenomenon is simply a restatement of the phenomenon, and when your “model” doesn’t actually specify the moving parts that help constrain your expectation, when your model implicitly assumes a separate solution to the problem, you don’t have a theory; you have cleverly-disguised ignorance.
when you seek a mate, the [control] reference [being tracked] is, of course, having a mate. You perceive that you do not have one, and take such steps as you think appropriate to find one.
You’re asking RichardKennaway to recapitulate a very complex physical description in a short comment. But “Behavior: The Control Of Perception” actually describes almost a dozen different modeling layers to get from simple intensity perception all the way up to high-level concepts and long-term goals, each building on the last, and taking 18 chapters to do it. The inferential gap between non-PCT thinking and PCT-thinking is way too big for condensing to a comment. Try arguing evolution to a group of creationists, and see how far you can get in one comment, without being accused of handwaving and tautology!
Hell, I just finished writing a primer on PCT-thinking for my subscribers, specifically regarding self-help applications: it’s 30 pages long. And I skipped all the meat of PCT’s neural-level predictions, the math, the experimental results, etc.
There’s a better (more detailed) way to explain mate tracking in PCT, that’s covered in the B:CP book, and it applies to all instinctual appetites. It’s called “intrinsic error”, and I’m a bit surprised Richard hasn’t mentioned it.
I’m thinking that maybe one difference between me and Richard, though, is that he’s a robotics researcher, and therefore very focused on the lower rungs of the PCT hierarchy. I’m interested in people, so I’m focused on the higher rungs, and how the rungs connect, especially to wired-in intrinsics. The elegant bit for me is that the intrinsic error model shows how easy it is to get (evolutionarily speaking) from instinctual behavior to controlled behavior to intelligent behavior, in incremental steps. (It also shows a likely connection to Ainslie’s conditioned-appetites model.)
But I’m not gonna sit down and rewrite a chapter of B:CP here in this comment to explain the bloody thing, because your inevitable follow-on questions will then require me to recapitulate the prior 10 or 11 chapters of B:CP, as well.
Je n’ai fait celle-ci plus longue que parce que je n’ai pas eu le loisir de la faire plus courte.
...which probably didn’t help your efforts, because translations vary and are often reworded slightly to suit the tastes of the quoter.
But I recalled that it was about a letter, and googling on letter shorter longer was sufficient to find the attribution. Perhaps my mind was already well-primed for finding relevant quotes and this illustrates well why I feel almost palpably stupider without internet access.
I misremembered the quote as involving an apology for the length, so my searches focused on the word “apologize”. But it didn’t matter in the end—you were my external brain. ;-)
If this is the problem, that’s fine (and I apologize for not giving more emphasis to your posts in replying on this issue). Inferential distance is a real problem.
But that’s not how Richard responded. He literally restated the problem in different terminology, replacing the problems with black boxes that have the solution inside. The only thing he said remotely related to inferential distance was that gosh, it’s hard to build up all the way from the neuron level.
Of course it is! That’s why AI is hard! But he’s trying to show one way that it simplifes by using an approach, which would mean he can do more than restate the problem. Even if the inferential distance is great, you can pain tbroad strokes, which would help identify where the disagreement is.
Here’s the nagging issue for me: I suspect that what’s going on is that you’ve come up with ad hoc techniques that happen to work, and then “explaining” them by a superficial resemblance to an actual controls problem. But “seeking a mate” actually describes a very complex set of behaviors, and it just doesn’t help to reframe that as “tracking the reference of ‘having a mate’ by outputting behaviors dependent on my distance (measured how?) from that state”.
Another data point for my claim is that you didn’t seriously approach the challenge I gave you, to check if something known not to work, would be deemed by PCT to work. That would require you to give an example and show where it parts with PCT, which is a pretty simple task.
Also you consider it to be a good thing when a theory requires you to separately solve the very problem it attacks, in order to use it. That suggests another level of confusion.
[me]By the time you’ve actually described what reference someone is tracking (or even a sub-reference like “sexiness”) and how observations are converted into a format capable of being compared, you’ve already solved the problem
[you]Yes, and that’s precisely what’s useful.
ETA: Compare that to:
By the time you’ve come up with all the epicyles needed to predict planet locations, you’ve have to already know the planet locations!
Yes, and that’s precisely why the Ptolemiac model is useful!
But that’s not how Richard responded. He literally restated the problem in different terminology, replacing the problems with black boxes that have the solution inside
I was being flippant. I mean, what were you expecting? Imagine that the person who first had the idea that thinking is done by neurons has just published it, and you ask them what you asked me. What can he tell you about finding a girlfriend? Only that it’s done by neurons. The leg work to discover just how those neurons are organised to do it is the problem, and finding a mate isn’t the place to start, experiments like Hubel and Wiesel’s on cat vision are the place to start, or mapping the nervous system of C. elegans.
Likewise, I’m not passing off “it’s done by control systems” as the solution of a problem, but as a basic insight that gives the beginning of a way to study living organisms. The place to begin that study and establish exactly what control systems are present and how they work is in studies like the one that you dismissed as a trivial game.
That’s what real work looks like. Have a spade. A start has been made at the various PCT links I’ve posted. Maybe in 50 years you’ll get an answer. But don’t be downhearted—it’s been more than a century so far for “it’s made of neurons”.
The place to begin that study and establish exactly what control systems are present and how they work is in studies like the one that you dismissed as a trivial game.
Telling a person “Perform this task, which involves acting like a control system” and discovering that people can, indeed, act like a control system doesn’t seem to demonstrate that people are physically made out of control systems. My desktop computer isn’t a control system, as such, but I can emulate a crude thermostat with a few lines of pseudocode...
The person performing that task is not “acting like” a control system, they actually are controlling the prescribed variable. The hypothesis is that living organisms are, in fact, constituted in this manner, with many control systems in a particular hierarchical arrangement. That every action they perform is an output action of a control system that is endeavouring to keep some perception at some reference level.
But I’ve belaboured this enough in this thread. Any more would just be a repetition of the materials I’ve pointed to.
The person performing that task is not “acting like” a control system, they actually are controlling the prescribed variable. The hypothesis is that living organisms are, in fact, constituted in this manner, with many control systems in a particular hierarchical arrangement. That every action they perform is an output action of a control system that is endeavouring to keep some perception at some reference level.
Indeed. I don’t disagree with anything here.
What I’m trying to say is that the ability to control one variable doesn’t provide much evidence for “it’s control systems all the way down”. One might as well claim “The brain is a finite state machine” because we can simulate them using pencil and paper.
Such modesty! It’s actually worse than that. You could write a program for a feedforward thermostat (i.e. which tries to predict how much to heat or cool based on factors other than the room temperature, like the sunshine, temp outside, insulation, etc.) on your computer, but Powers et al. would scream bloody murder if you tried to use that as evidence that living systems are feedforward control loops!
You could write a program for a feedforward thermostat
Actually, you couldn’t. At least, it wouldn’t work very well, not nearly as well as a system that simply measures the actual temperature and raises or lowers it as necessary.
Try it and see.
“Feedforward control loop” is pretty much a contradiction in terms. Look at anything described as feedforward control, and you’ll find that it’s wrapped inside a feedback loop, even if only a human operator keeping the feedforward system properly tuned. There are some demonstrable feedforward links in the nervous system, such as the vestibulo-ocular reflex, but as expected, the VOR is wrapped inside a feedback system that tunes its parameters. It wouldn’t work without that.
Actually, you couldn’t. At least, it wouldn’t work very well, not nearly as well as a system that simply measures the actual temperature and raises or lowers it as necessary.
Ah, but if I deliberately created an artificial scenario designed to make FF control work, then FF control would look rockin’.
You know, like the programs you linked do, except that they pimp feedback instead ;-)
Yes, feedback control is usually better; my point was the excessive extrapolation from that program.
“Feedforward control loop” is pretty much a contradiction in terms. Look at anything described as feedforward control, and you’ll find that it’s wrapped inside a feedback loop, even if only a human operator keeping the feedforward system properly tuned.
Yes, very true, which reminds me: I saw a point in the demo1 program (link when I get a chance) on the site pjeby linked where they have you try to control a system using either a) your knowledge of the disturbance (feedforward), or b) your knowledge of the error (feedback), and you inevitably do better with b).
Here’s the thing though: it noted that you can get really good at a) if you practice it and get a good feel for how the disturbance relates to how you should move the mouse. BUT it didn’t use this excellent opportunity to point out that even then, such improvement is itself due to another feedback loop! Specifically, one that takes past performace as the feedback, and desired performance as the reference.
my point was the excessive extrapolation from that program.
PCT is not derived from the demos; the demos are derived from PCT.
even then, such improvement is itself due to another feedback loop
So you see, wherever you look in the behaviour of living organisms, you find feedback control!
If that seems trivial to you, then it is probably because you are not an experimental psychologist, which is the area in most need of the insight that living organisms control their perception. You probably also do not work in AI, most of whose practitioners (of strong or weak AI) are using such things as reinforcement learning, planning, modelling, and so on. Robotics engineers—some of them—are about the only exception, and they have a better track record of making things that work.
BTW, I’m not touting PCT or anything else as the secret of real AI. Any better understanding of how real brains operate, whether it comes from PCT or anything else, will presumably facilitate making artificial ones, and I have used it as the basis of a fairly good (but only simulated) walking robot, but strong AI is not my mission.
In that case, why do you keep insinuating that control theory is useful as a high-level model of intelligence? That seems analogous to deriding computational neuroscientists for not using neurons in their models.
ETA: By comparison, evolutionary psychologists can’t do the math yet on the selective advantage of genes coding for variations in mental traits, but their derived models of human psychology allow them to make significant predictions that weren’t thought of without the model, and which often check out in experiments. Does PCT have any novel experimental consequences that have been verified?
I was being flippant. I mean, what were you expecting?
I was expecting you to show me how the controls paradigm does a better job of explaining behaviors than the prevailing, but messy, model of evolutionary psychology. (Not that they would contradict, but your model would have to be simpler and/or more precise.)
That was the challenge I had presented to you before: show how “humans are control systems” is better at compressing a description of our observations.
If it seems hard to do, it’s probably because it is hard to, because it’s not a better model, because the behavior is so hard to express in controls terminology. Finding a mate is simply not like making sure one line is under another.
Imagine that the person who first had the idea that thinking is done by neurons has just published it, and you ask them what you asked me.
If I were the first person to publish the neuron theory, it would include an actual model with actual moving parts and therefore have actual explanatory power over actual observations that actual other people can make. It would not say, as you have essentially done, “people think with neurons, so like, when you think, it’s all … neurony. Are you thinking about sex? Yeah, the neuronal model has that. Neurons cause that thinking too. See, once I know what you’re thinking about, I can predict, using the neuronal model, what you’re thinking.”
The leg work to discover just how those neurons are organised to do it is the problem, and finding a mate isn’t the place to start,
But human behavior is where you’ve started, hence my confusion of what exactly the “humans as controllers” model accomplishes.
Likewise, I’m not passing off “it’s done by control systems” as the solution of a problem, but as a basic insight that gives the beginning of a way to study living organisms.
But it’s not an insight unless it makes the problem easier. Everything you’ve presented here simply shows how you could rephrase the solution to predicting organism behavior into a controls model once it’s been solved. So where does the ease come in? What problem becomes easier when I approach it your way?
The place to begin that study and establish exactly what control systems are present and how they work is in studies like the one that you dismissed as a trivial game.
I dismissed it as a trivial game because it is a trivial game. From the fact that I use proportional feedback control to keep two lines together, it does not follow that this is a useful general model of all organism activity.
Believe it or not, I have put a lot of work (in terms of fraction of my spare time) into exploring PCT. I’ve completed demo1 on the site pjeby linked, and have run it and four others in DOS Box. I’ve read some of the pdfs expalining feedback systems at the cellular level. I am going far out of my way to see if there’s anything to PCT, so please do not write me off as if I’m making you hold my hand.
I am making every effort to make things easy for you. All that’s left for your is to have a real model that you actually understand.
That, in turn, would rebut my strongly justified suspicion that the model’s “predictions” are ad hoc and the parallels with control systems superficial.
The reason why expressing the connection between not having a mate and seeking a mate in terms of PCT is so difficult is because “not having a mate” is not a perception, and because “seeking a mate” is not a behavior. Rather, these are an abstract world state with multiple perceptual correlates, and a broad class of complex behaviors that no known model explains fully. Given such a confusing problem statement, what did you expect if not a confused response?
The second problem, I think, is that you may have gotten a somewhat confused idea of what (non-perceptual) control systems look like. There was a series of articles about them on LW, but unfortunately, it stopped just short of the key insight, which is the PID controller model. A PID controller looks at not just the current value of its sensor (position, P), but also its recent history (integral, I) and rate of change (derivative, D).
If you want to test PCT, you need to step back and look at something simpler. The most obvious example is motor control. Most basic motor control tasks, like balancing, are a matter of generating some representation of body and object position, figuring out which neurons trigger muscles to push it in certain ways, and holding position constant; and to do that, any organism, whether it’s a human or a simple invertebrate, needs some neural mechanism that acts very much like a PID controller. That establishes that controllers are handled in neurology somehow, but not their scope. There’s another example, however, which shows that it’s considerably broader than just motor control.
Humans and animals have various neurons which respond to aspects of their biochemistry, such as concentrations of certain nutrients and proteins in the blood. If these start changing suddenly, we feel sick and the body takes appropriate action. But the interesting thing is that small displacements which indicate dietary deficiencies somehow trigger cravings for foods with the appropriate nutrient. The only plausible mechanism I can think of for this is that the brain remembers the effect that foods had, and looks for foods which displaced sensors in the direction opposite the current displacement. The alternative would be a separate chemical pathway for monitoring each and every nutrient, which would break every time the organism became dependent on a new nutrient or lost access to an old one.
Moving up to higher levels of consciousness, things get significantly more muddled. Psychology and clear explanations have always been mutually exclusive, and no single mechanism can possibly cover everything, but then it doesn’t need to, since the brain has many obviously-different specialized structures within it, each of which presumably requires its own theory. But I think control theory does a good job explaining a broad enough range of psychological phenomena that it should be kept in mind when approaching new phenomena.
Moving up to higher levels of consciousness, things get significantly more muddled.
I disagree, but that’s probably because I’ve seized on PCT as a compressed version of things that were already in my models, as disconnected observations. (Like time-delayed “giving up” or “symptom substitution”.) I don’t really see many gaps in PCT because those gaps are already filled (for me at least), by Ainslie’s “conditioned appetites” and Hawkins’ HTM model.
Ainslie’s “interests” model is a very strong fit with PCT, as are the hierarchy, sequence, memory, and imagination aspects of HTM. Interests/appetites and HTM look just like more fleshed-out versions of what PCT says about those things.
Is it a complete model of intelligence and humans? Heck no. Does it go a long way towards reverse-engineering and mapping the probable implementation of huge chunks of our behavior? You bet.
What’s still mostly missing, IMO, after you put Ainslie, PCT, and HTM together, is dealing with “System 2” thinking in humans: i.e. dealing with logic, reasoning, complex verbalizations, and some other things like that. From my POV, though, these are the least interesting parts of modeling a human, because these are the parts that generally have the least actual impact on their behavior. ;-)
So, there is little indication as to whether System 2 thinking can be modeled as a controller hierarchy in itself, but it’s also pretty plain that it is subject to the System 1 control hierarchy, that lets us know (for example) whether it’s time for us to speak, how loud we’re speaking, what it would be polite to say, whether someone is attacking our point of view, etc. etc.
It’s also likely that the reason we intuitively see the world in terms of actions and events rather than controlled variables is simply because it’s easier to model discrete sequences in a control hierarchy, than it is to directly model a control hierarchy in another control hierarchy! Discrete symbolic processing on invariants lets us reuse the controllers representing “events”, without having to devote duplicated circuitry to model other creatures’ controller hierarchies. (The HTM model has a better detailed explanation of this symbolic/pattern/sequence processing, IMO, than PCT, even though in the broad strokes, they’re basically the same.)
(And although you could argue that the fact we use symbols means they’re more “compressed” than control networks, it’s important to note that this is a deliberately lossy compression; discrete modeling of continuous actions makes thinking simpler, but increases prediction errors.)
The reason why expressing the connection between not having a mate and seeking a mate in terms of PCT is so difficult is because “not having a mate” is not a perception, and because “seeking a mate” is not a behavior. Rather, these are an abstract world state with multiple perceptual correlates, and a broad class of complex behaviors that no known model explains fully. Given such a confusing problem statement, what did you expect if not a confused response?
What a pitiful excuse.
Let’s get some perspective here: the model I’m trying to understand is so vague in the first place (in terms of what insight it has to offer), despite all of my efforts to understand it with basic questions about what it does to replace existing models. Of course my questions about such an ill-supported paradigm are going to look confused, but then again, it’s not my responsibility to make the paradigm clear. That burden, currently unmet, lies on the person presenting it.
If you’re familiar with the tools of rationality, it is a trivial task to handle “confused” questions of exactly the kind I just asked—but that pre-supposes you have a clue what you’re talking about. All you have to do is identify the error it makes, find the nearest meaningful problem, and show how your model handles that.
A confused response is neither appropriate, nor deserved, and only reflects poorly on the responder.
Let me show you how it works. Let’s say I’m some noble defender of the novel Galilean model, trying to enlighten the stubborn Ptolemaic system supporters. Then, some idiot comes along and asks me, “Okay, smarty, how does the Galilean model plot the epicycle for Jupiter?”
In response, do I roll my eyes at how he didn’t use my model’s terminology and hasn’t yet appreciated my models beauty? Do I resign myself to giving a “confused response”? No.
And do you know why I don’t? Because I have an actual scientific model, that I actually understand.
So in response to such a “hopeless” dilemma, I marshal my rationalist skills and give a non-confused response.
Ready to have your mind blown? Here goes:
“When you ask me about Jupiter’s epicycle, what you’re really looking for is how to plot its position relative to earth. But my point is, you don’t need to take this step of calculating or looking up epicycles. Rather, just model Jupiter as going around the sun in this well-defined eilliptical path, and the earth in this other one. We know where they will be relative to the sun as a function of time, so finding Jupiter relative to the earth is just matter of adding the earth-to-sun vector to the sun-to-jupiter vector.”
There, that wasn’t so hard, was it? But, I had it easy in that I’m defending an actual model that actually compresses actual observations. Richard, OTOH, isn’t so lucky.
Notice what I did not say: “You find Jupiter in the sky and then you draw an epicycle consistent with its position, but with the earth going around the sun”, which is about what I got from Richard.
Moving up to higher levels of consciousness, things get significantly more muddled.
Yeah, that’s the point. Those higher levels are exactly what pjeby attempts to use PCT for, which is where I think any usefulness (of the kind seen in biochemical feedback loops) loses its compression abilities, and any apparent similarity to simple feedback control systems is superficial and ad-hoc, which is exactly why no one seems to be able to even describe the form of the relationship between the higher and lower levels in a way that gives insight. That is, break down “finding a mate” into related controllable values and identify related outputs. Some specification is certainly possible here, no?
Given such a confusing problem statement, what did you expect if not a confused response?
What a pitiful excuse.
[...]
A confused response is neither appropriate, nor deserved, and only reflects poorly on the responder.
Neither the problem statement, nor any of the confused responses were mine. My post was meant to clarify, not to excuse anything.
If you’re familiar with the tools of rationality, it is a trivial task to handle “confused” questions of exactly the kind I just asked—but that pre-supposes you have a clue what you’re talking about. All you have to do is identify the error it makes, find the nearest meaningful problem, and show how your model handles that.
No, that is not the correct way to handle confused questions. The correct way to handle them is to back up, and explain the issues that lead to the confusion. In this case, there are many different directions the question could have been rounded in, each of which would take a fairly lengthy amount of text to handle, and people aren’t willing to do that when you could just say that wasn’t what you meant. I should also observe that pjeby gave you a citation and ducked out of the conversation, specifically citing length as the problem.
At some point, you seem to have switched from conducting a discussion to conducting a battle. Most of the parent post is not talking about the supposed topic of discussion, but about the people who participated in it before. Unfortunately, the history of this thread is far too long for me to read through, so I cannot respond to those parts. However, I am strongly tempted to disregard your arguments solely on the basis of your tone; it leads me to believe that you’re in an affective death spiral.
Neither the problem statement, nor any of the confused responses were mine.
I know. Still a pitiful excuse, and yes, it was an excuse; you insinuated that my confused question deserved the flippant response. It didn’t. It required a simple, clear answer, which of course can only be given when the other party actually has a model he understands.
No, that is not the correct way to handle confused questions. The correct way to handle them is to back up, and explain the issues that lead to the confusion.
We’re bickering over semantics. The point is, there are more helpful answers, which one can reasonably be expected to give, than the “confused reply” you referred to. Richard knows what “finding a mate” means. So, if he actually understands his own model, he can break down “finding a mate” into its constituent references and outputs.
Or say how finding a mate should really be viewed as a set of other, specific references being tracked.
Or somehow give a hint that he understands his own model and can apply it to standard problems.
Was my epicycle example not the kind of response I could reasonably expect from someone who understands his own model?
But “seeking a mate” actually describes a very complex set of behaviors, and it just doesn’t help to reframe that as “tracking the reference of ‘having a mate’ by outputting behaviors dependent on my distance (measured how?) from that state”.
That’s because, as I said, Richard completely handwaved the PCT explanation, in the same way as an evolution supporter would likely end up handwaving ideas like “inclusive genetic fitness”.
To model mate searching in PCT, you would need to include several continuous variables related to mate selection, such as “how attractive am I” and “how often am I getting laid”, that would be used to control other variables like for “minimum required desirability-of-partner”. Lumping that all together into one variable as Richard suggested would be quite ludicrous. (But note, by the way, that there’s mainstream evidence for the existence of such variables in humans and other animals.)
Keep in mind that controllers are hierarchical, so low-level behavior patterns are driven by higher level ones. So, the “minimum partner desirability” threshold would gate the program-level controller for mate attraction behaviors… which would include various levels to track of the potential mate’s apparent response or interest, etc. You’ve got levels going all the way down here, implied by “hierarchical control”, in the same way that a huge host of behaviors and characteristcs are covered by “inclusive genetic fitness”.
Another data point for my claim is that you didn’t seriously approach the challenge I gave you, to check if something known not to work, would be deemed by PCT to work. That would require you to give an example and show where it parts with PCT, which is a pretty simple task.
I did that, in precisely the comment you linked to as saying I didn’t. Specifically, I pointed out that the exact set of definitions for e.g. “thinking hard” that wouldn’t work in reality, precisely match the ones that PCT predicts would not work, and vice versa for the definitions that would not work.
If you’re saying, you don’t see how that’s so from what I wrote and need more detail, that’s fine. But to me, I did exactly as you requested, or the best I could given your undefined phrase “thinking hard”.
Also you consider it to be a good thing when a theory requires you to separately solve the very problem it attacks, in order to use it.
You quoted me out of context; the rest of the comment goes on to explain that what I find useful is that PCT tells us what things to look for in order to solve the problem.
What I perhaps didn’t explain very well is how those “things to look for” differ from what might otherwise be looked for. For example, PCT emphasizes integrated continuous variables rather than discrete events, e.g. “probability of receiving a shock within time period t”. Our normal thinking about behavior emphasizes discrete goals and actions, while one of PCT’s central ideas is that discrete goals and actions occur only to restore continuous variables to their reference ranges, in response to either environmental disruptions, or the passage of time (causing an integrated or “average” level to drop).
That’s because, as I said, Richard completely handwaved the PCT explanation, in the same way as an evolution supporter would likely end up handwaving ideas like “inclusive genetic fitness”.
No, not like an evolution supporter, because an evolution supporter could identify what exactly IGF refers to in a way that is not a trivial restatement of the problem of “Why would a mother give up her life for her two children?”
A scientific-sounding answer would be, “because that improves her IGF”, and you’re correct this would be a handwave.
A really scientific answer would “because that improves her IGF, which is roughly the fraction of the next generation’s genes that are hers, which would account for why the gene pool is dominated by the genes of people who made such a tradeoff”.
Richard does not have such a moving-parts model that breaks down the concepts of “having a mate” and “recognition of having a mate” and “action that moves me toward having a mate”. They’re irreducible black boxes, at least as far as he’s aware. That’s not a theory. If he had a theory, he should have and would have focused more on expanding one of these concepts than on making sure the problem was fully restated with with different terms.
To model mate searching in PCT, you would need to include several continuous variables …
in other words, give an answer different from RIchard. Hint: if the very same model can be “interpreted” to imply different explanations, it’s not a good model nor a theory.
That would require you to give an example and show where it parts with PCT, which is a pretty simple task.
I did that, in precisely the comment you linked to as saying I didn’t. Specifically, I pointed out that the exact set of definitions for e.g. “thinking hard” that wouldn’t work in reality, precisely match the ones that PCT predicts would not work, and vice versa for the definitions that would not work.
No, you asserted it, with no reference to the specific aspects of PCT that rule against the “non-phenomena”.
Surely my query warranted more of an answer than, “yep, everything’s okay over here” even if you went through the effort to transform it to, “yep, what I’ve seen work, it sure is expected to work by PCT, and what I know doesn’t work, yep, it sure is claimed by PCT not to work!” (Edit: removed some snarkiness)
Also you consider it to be a good thing when a theory requires you to separately solve the very problem it attacks, in order to use it.
You quoted me out of context; the rest of the comment goes on to explain that what I find useful is that PCT tells us what things to look for in order to solve the problem.
The context makes you look worse! Here it is:
Yes, and that’s precisely what’s useful. That is, it identifies that to solve anyone’s problems, you need only identify the reference values, and find a way to reorganize the control system to either set new reference values or have another behavior that changes the outside world to cause the new reference to be reached.
So basically, “all” you have to do is find the reference values and appropriately modify them … but specifying the reference values itself contains the exact same insights you’d need to solve the problem anyway! (it’s “AI-complete” in compsci/AI jargon) So the “human as controller” model doesn’t simplify the problem, it just says “here, go solve the problem, somehow, and when you do, without the help of this model, you’ll see that one of the six trillion neat things you can do is specify them in controls format”.
The fact that you’re still ignoring any of the substantive and responsive portions of my comments, bodes ill for this being a useful exchange.
It is quite possible I’ve misunderstood your queries and/or answered them inadequately. However, I’d like to think that the appropriate response in that case would be to clarify what you want, rather than simply taking it to mean no-one can give you what you want.
So the “human as controller” model doesn’t simplify the problem, it just says “here, go solve the problem, somehow, and when you do, without the help of this model, you’ll see that one of the six trillion neat things you can do is specify them in controls format”.
I’ve mentioned a number of things that PCT does beyond that. For example, it shows that the first thing to look for in modeling are continuous analog variables integrated over a time period, with shorter time periods generally being represented lower in the control hierarchy than longer time periods. AFAICT, this is far from an obvious or trivial modeling distinction.
The fact that you chose not to comment on that, but instead dug in on justifying your initial position, suggests to me that you aren’t actually interested in the merits (or lack thereof) of PCT as a modeling tool, so much as in defending your position.
The fact that you’re still ignoring any of the substantive and responsive portions of my comments, bodes ill for this being a useful exchange.
Yeah, I like that strategy. “In this extremely long, involved exchange, any part of my post that you didn’t directly respond to, was an ultra-critical omission, and completely demonstrates your failure to act in good faith or adequately respond to my points.”
Whatever. I didn’t respond to it because I haven’t gotten around to responding to your points in the other thread, or because it didn’t address my request. In this case, it’s the latter.
For example, it shows that the first thing to look for in modeling are continuous analog variables integrated over a time period, with shorter time periods generally being represented lower in the control hierarchy than longer time periods. AFAICT, this is far from an obvious or trivial modeling distinction.
Okay, and what epistemic profit does this approach gain for you, especially given that deliberate actions in pursuit of a goal are highly discontinuous? Oh, right, add another epicycle. Hey, the Hawkins HTM model, that’ll work!
ETA: Do not interpret this post to mean I’m in full anti-PCT mode. I am still exploring the software on the site pjeby linked and working through the downloadable pdfs. I’m making every effort to give PCT a fair shake.
especially given that deliberate actions in pursuit of a goal are highly discontinuous?
I’m not certain I understand your terms. If I interpret your words “classically”, then of course I “know what you mean”. However, if I’m viewing them through the PCT lens, those words make no sense at all, or are blatantly false.
When you drive a car and step on the brake, is that a “deliberate action” that’s “discontinuous”? Classically, it seems obvious. PCT-wise, you’re begging the question.
From the PCT perspective, the so-called “action” of braking is a chain of controls looking something like:
Speed-change controller detects discrepancy between current acceleration and desired deceleration, sets braking controller to “braking hard”
Braking controller notes we aren’t braking, sets foot position to “on brake”
Foot position controller detects foot is out of position, requests new leg position
Leg position controller detects out of position, requests new leg speed/direction
Leg speed controller detects not moving, requests increased muscle force
...etc., until
Foot position controller detects approaching correct position, and lowers requested movement speed, until desired position is reached
Speed controller observes drop of speed below its reference level, sets speed-change controller to “slow accelerate”
Speed-change controller notices that current deceleration is below “slow accelerate” reference, sets “gas” controller to “slight acceleration”
...and so on, until speed stabilizes… and the foot goes up and down slightly on the gas… all very continuously.
So, there is nothing at all “discontinuous” about this. (Modulo the part where nerves effectively use pulse-width modulation to communicate “analog” values).
And it’s precisely this stable continuity of design that makes PCT so elegant; it requires very little coordination (except hierarchically), and it scales beautifully, in the sense that mostly-identical control units can be used. Got a more complex animal? Need more sophisticated behavior? Just add controllers, or new layers of controllers.
Need a new skill? Learning grows in or assigns some new controllers, that measure derived perceptual quantities like “speed of the car”, “braking”, and “putting on the gas”. (Which explains why procedural knowledge is more persistent than propositional knowledge—the controllers represent a hardware investment in knowledge.)
And within this model, actions are merely side-effects of disturbances to the regulated levels of perceptual variables, such as speed. I stopped the upward point of the hierarchy at the speed controller noticing a speed discrepancy, but the reason for that discrepancy could be you noticing you’re late, or it could be that your “distance to next car” controller has issued a request to set the new “desired speed” to “less than the car in front of us”. In either case, the “action” is the same, regardless of what “goal”—or more likely, disturbance—caused it to occur.
That being said, PCT does include “sequence” and “program” controller layers, that can handle doing things in a particular sequence or branching. However, even these are modeled in terms of a perceptual control hierarchy, ala TOTE loops. That is, you can build TOTE loops by wiring controllers together in relatively simple ways.
Reification of programs and “actions” through controller hierararchies is also a good strategy for building a fast machine out of slow components. Rather than share a few ultra-fast, complex components, PCT hierarchies depend on chains of similar, simultaneously-responding, cheap/dumb components, such that the fastest responses are required from the components that are generally nearest (network-wise) to the place where the signals need to be received or delivered to exert control.
These are just some of the obvious properties that make PCT-style design a good set of tradeoffs for designing living creatures, using similar constraints to evolution. (Such as the need to be able to start with primitive versions of the model, and gradually scale up from there.)
Okay, and what epistemic profit does this approach gain for you
As I said, it gives me a better idea of what to look for. After grasping PCT, I was able to identify certain “bugs” in my brain that had previously been more elusive. The time and hierarchy distinctions made it possible for me to identify what I was controlling for, rather than just looking at discrete action triggers, as I did in the past.
In this area, PCT provides a more compact model of what psychologists call “secondary gain” , hypnosis people call “symptom conversion”, and NLP people call “ecology”.
The idea is that when you take away one path for someone to get something (e.g. giving up smoking) they may end up doing something else to satisfy a need that was previously supported by the old behavior (e.g. chewing gum).
What psychologists, NLPers, and hypnosis people never had a good explanation for (AFAIK) is why it takes time for this substitution or reversion to occur! Similarly, why does it take time for people to stop persisting at trying to do something new?
This is an example of a complex behavioral property of humans that falls directly out of the PCT model without any specific attempt to generate it. Since high-level goals are integrated over a longer time period, it takes time for the error signal to rise, and then further time for the controller network reorganization process (part of the PCT model of learning) to find an alternative or extinguish the changed behavior.
I find PCT parsimonious because there are so many little quirks of human nature I know about, that would be naturally expected to occur if behavior was control-system driven in precisely the ways PCT predicts that it is… but which are just weird and/or unexplained under any other model that I know of.
From the PCT perspective, the so-called “action” of braking is a chain of controls looking something like: [...]
Okay, thank you, that was exactly the kind of answer I was looking for, in terms of breaking down (what is framed by us non-PCTers as) a discrete list of actions into hierarchical feedback loops and what they’re using for comparison. Much appreciated.
But just the same, I think your explanation illuminates my complaint about the usefulness of the model. What it appears to me is, you just took a list of discrete steps and rephrased them as continuous values. So far, so good, but all I see is added complexity. Let me explain.
I would describe my steps in baking a cake (and of course this abstracts away from lower level detail) as:
1) Open preheated oven.
2) Place pan containing batter onto middle of middle over rack.
3) Close oven.
4) Set timer.
Your claimed improvement over this framing of these events is:
1) Define variable for oven openness. Recognize it’s zero and push it toward 1.
2) Define variable for pan distance from middle of middle oven rack. Recognize it’s too high and push it toward zero.
3) Recognize oven openness is 1 and should be zero, push it in that direction.
4) Define variable for oven-timer-value-appropriateness. Recognize it’s too low and move it higher.
Yes, superficially, you’ve made it all continuous, but only by positing new features of the model, like some neural mechanism isomorphic to “detection of oven-timer-value-appropriateness”, which requires you to expand that out into another complex mechanism.
I agree, as I’ve said before, that this is one way to rephrase what is going on. But it doesn’t simplify the problem; it forces you identify the physical correlate of “making sure the oven’s set to the right time” in a form that I’m not convinced is appropriate for the problem. Why isn’t it appropriate?
Among other things, you’re forced to solve the object recognition problem and identify a format for comparison. But if I’ve solved the (biological) object recognition problem, my model can simply invoke the actual neural mechanism being used, without the added complexity of reformatting the causal flow into feedback loops.
You defend this model by its elegance, but you only get the elegance after you solve the problem some other way. That is, I only have an elegant hierarchical feedback loop if I can, somehow, solve the object recognition problem that allows me to actually specify a reference and feedback signal. A model isn’t any good if it presupposes the solution of the problem it’s being used to solve.
I would describe my steps in baking a cake (and of course this abstracts away from lower level detail)
You could describe them that way, yes, and that would nominally describe the behaviors you emit. However, it’s trivial to prove that this does not describe the implementation in your head that emits those behaviors!
For example, you might forget to preheat the oven, in which case the order of your steps is going to change. There are any number of disruptions that can occur in your sequence of “steps”, that will cause you to change your actions to work around the disruptions, with varying degrees of automaticity, depending on how high in your control hierarchy the disruptions reach.
A simple disruption like a spill on the floor you need to walk around will be handled with barely a conscious notice, while a complex disruption like the power being off (and the oven therefore not working) will induce more complex behavior requiring conscious attention.
If a sequence of steps could actually describe human behavior, we could feed your list of steps to a computer and get it done. It’s actually omitting some of the most important information: the goals of the steps, and how to measure whether they’ve been obtained.
And that’s information our brains have to be able to know and use in order to actually carry out behavior. We tend to assume we do this by “thinking”, but we usually only “think” in order to handle high-level disturbances that require rearrangement of goals, rather than just using existing control systems to work around the disturbance.
Your claimed improvement over this framing of these events is: [list I wouldn’t use to describe it]
When you get to higher levels of modeling, you can certainly deal with sequences of subgoals that are matching of perceptual patterns like “cake is in the oven”. Did you read the TOTE loops reference I gave? TOTE loops act on something until it reaches a certain state, then activate another TOTE loop. PCT incorporates the previously-proposed cog psych notion of TOTE loops, and proposes some ways to build TOTE loops out of simpler controllers.
Part of the modeling elegance of using controllers and TOTE loops to implement overall behavioral programs is that they allow you to notice, for example, that your assistant chef has already set the timer for you as you placed the cake in the oven… and thereby skipping the need to perform that step.
Among other things, you’re forced to solve the object recognition problem and identify a format for comparison. But if I’ve solved the (biological) object recognition problem, my model can simply invoke the actual neural mechanism being used, without the added complexity of reformatting the causal flow into feedback loops.
This I don’t get. We already know that (visual) object recognition can be implemented by time-varying hierarchical feature sequences tied to the so-called “grandmother neurons”. Both HTM and PCT include this concept, except that PCT proposes you get an analog “grandmotherness” signal, whereas IIRC the HTM model assumes it’s a digital “grandmother present” signal. But HTM at least has pattern recognition demos that handle automatic learning and pattern extraction from noisy inputs, and it uses exactly the same sort of recognition hierarchy that the full PCT model calls for.
That’s why I keep saying that if you want to see the entire PCT model, you need to read the book. Most of the primers either talk about low-level stuff or high-level stuff. Object recognition, action sequences, and that sort of thing are all in the middle layers that make up the bulk of the book.
You defend this model by its elegance, but you only get the elegance after you solve the problem some other way. That is, I only have an elegant hierarchical feedback loop if I can, somehow, solve the object recognition problem that allows me to actually specify a reference and feedback signal. A model isn’t any good if it presupposes the solution of the problem it’s being used to solve.
Note, by the way, that this is backwards, when applied to listing steps and calling it a model. The steps cannot be used to actually predict behavior, because they list only the nominal case, where everything goes according to plan. The extra information that PCT forces you to include results in a more accurate model—one that does not simply elide or handwave away the parts of behavior that we intuitively ignore.
That is, the parts we don’t usually bother communicating to other human beings, because we assume they’ll fill in the gaps.
PCT is useful because it shows where those gaps are and what is needed to fill them in, in much the same way that the initial description of evolution identified gaps in our knowledge about biology, and what information was needed to fill them in, in place of the asumptive handwaving that “God did it”. In the same way, we currently handwave most of what we don’t understand about behavior as, “X did it”, where X is some blurry entity or other such as learning, environment, intelligence, habit, genetics, conditioning, etc.
(And no, PCT doesn’t merely replace X with “control systems”, because it shows HOW control systems can “do it”, whereas other values of X are simply stopping the explanation at that point.)
Yeah, is it too much to ask that in a two-line top-level post, over half of which is transcribed to get stuff like that right?
Anyway, I’d like to self-servingly nominate this recent exchange as a case study in more advanced baloney detection, where someone’s supercool, unfairly suppressed Theory of Everything turns out to be, “Um, find the solution some other way, and then shoehorn it into my theory.”
sigh
It is not supercool, it is not suppressed, it is not a theory of everything, and there is no shoehorning. It is merely a theory of living organisms that works where it has been applied, and it has been applied.
But everyone, feel free to apply Shermer’s advice, or whatever more advanced techniques of rationality you have, to the matter. I hope for no less in posting here.
When your theory’s “explanation” of a phenomenon is simply a restatement of the phenomenon, and when your “model” doesn’t actually specify the moving parts that help constrain your expectation, when your model implicitly assumes a separate solution to the problem, you don’t have a theory; you have cleverly-disguised ignorance.
For example, when you say something like this:
You’re asking RichardKennaway to recapitulate a very complex physical description in a short comment. But “Behavior: The Control Of Perception” actually describes almost a dozen different modeling layers to get from simple intensity perception all the way up to high-level concepts and long-term goals, each building on the last, and taking 18 chapters to do it. The inferential gap between non-PCT thinking and PCT-thinking is way too big for condensing to a comment. Try arguing evolution to a group of creationists, and see how far you can get in one comment, without being accused of handwaving and tautology!
Hell, I just finished writing a primer on PCT-thinking for my subscribers, specifically regarding self-help applications: it’s 30 pages long. And I skipped all the meat of PCT’s neural-level predictions, the math, the experimental results, etc.
There’s a better (more detailed) way to explain mate tracking in PCT, that’s covered in the B:CP book, and it applies to all instinctual appetites. It’s called “intrinsic error”, and I’m a bit surprised Richard hasn’t mentioned it.
I’m thinking that maybe one difference between me and Richard, though, is that he’s a robotics researcher, and therefore very focused on the lower rungs of the PCT hierarchy. I’m interested in people, so I’m focused on the higher rungs, and how the rungs connect, especially to wired-in intrinsics. The elegant bit for me is that the intrinsic error model shows how easy it is to get (evolutionarily speaking) from instinctual behavior to controlled behavior to intelligent behavior, in incremental steps. (It also shows a likely connection to Ainslie’s conditioned-appetites model.)
But I’m not gonna sit down and rewrite a chapter of B:CP here in this comment to explain the bloody thing, because your inevitable follow-on questions will then require me to recapitulate the prior 10 or 11 chapters of B:CP, as well.
I am in awe of your productivity.
I think that for pjeby, writing a shorter primer would take longer...
For almost anyone who actually has something to communicate in writing, making it shorter will take longer.
-- Blaise Pascal
I wanted to use that quote, but I couldn’t recall the attribution or exact wording and my google-fu wasn’t up to finding it. Thanks!
Well, via Wikiquote the exact wording is:
...which probably didn’t help your efforts, because translations vary and are often reworded slightly to suit the tastes of the quoter.
But I recalled that it was about a letter, and googling on letter shorter longer was sufficient to find the attribution. Perhaps my mind was already well-primed for finding relevant quotes and this illustrates well why I feel almost palpably stupider without internet access.
I misremembered the quote as involving an apology for the length, so my searches focused on the word “apologize”. But it didn’t matter in the end—you were my external brain. ;-)
If this is the problem, that’s fine (and I apologize for not giving more emphasis to your posts in replying on this issue). Inferential distance is a real problem.
But that’s not how Richard responded. He literally restated the problem in different terminology, replacing the problems with black boxes that have the solution inside. The only thing he said remotely related to inferential distance was that gosh, it’s hard to build up all the way from the neuron level.
Of course it is! That’s why AI is hard! But he’s trying to show one way that it simplifes by using an approach, which would mean he can do more than restate the problem. Even if the inferential distance is great, you can pain tbroad strokes, which would help identify where the disagreement is.
Here’s the nagging issue for me: I suspect that what’s going on is that you’ve come up with ad hoc techniques that happen to work, and then “explaining” them by a superficial resemblance to an actual controls problem. But “seeking a mate” actually describes a very complex set of behaviors, and it just doesn’t help to reframe that as “tracking the reference of ‘having a mate’ by outputting behaviors dependent on my distance (measured how?) from that state”.
Another data point for my claim is that you didn’t seriously approach the challenge I gave you, to check if something known not to work, would be deemed by PCT to work. That would require you to give an example and show where it parts with PCT, which is a pretty simple task.
Also you consider it to be a good thing when a theory requires you to separately solve the very problem it attacks, in order to use it. That suggests another level of confusion.
ETA: Compare that to:
I was being flippant. I mean, what were you expecting? Imagine that the person who first had the idea that thinking is done by neurons has just published it, and you ask them what you asked me. What can he tell you about finding a girlfriend? Only that it’s done by neurons. The leg work to discover just how those neurons are organised to do it is the problem, and finding a mate isn’t the place to start, experiments like Hubel and Wiesel’s on cat vision are the place to start, or mapping the nervous system of C. elegans.
Likewise, I’m not passing off “it’s done by control systems” as the solution of a problem, but as a basic insight that gives the beginning of a way to study living organisms. The place to begin that study and establish exactly what control systems are present and how they work is in studies like the one that you dismissed as a trivial game.
That’s what real work looks like. Have a spade. A start has been made at the various PCT links I’ve posted. Maybe in 50 years you’ll get an answer. But don’t be downhearted—it’s been more than a century so far for “it’s made of neurons”.
Telling a person “Perform this task, which involves acting like a control system” and discovering that people can, indeed, act like a control system doesn’t seem to demonstrate that people are physically made out of control systems. My desktop computer isn’t a control system, as such, but I can emulate a crude thermostat with a few lines of pseudocode...
while(1) {
while(DesiredTemp > ActualTemp) {runAirConditioner(); }
while(DesiredTemp < ActualTemp) {runFurnace(); }
}
The person performing that task is not “acting like” a control system, they actually are controlling the prescribed variable. The hypothesis is that living organisms are, in fact, constituted in this manner, with many control systems in a particular hierarchical arrangement. That every action they perform is an output action of a control system that is endeavouring to keep some perception at some reference level.
But I’ve belaboured this enough in this thread. Any more would just be a repetition of the materials I’ve pointed to.
Indeed. I don’t disagree with anything here.
What I’m trying to say is that the ability to control one variable doesn’t provide much evidence for “it’s control systems all the way down”. One might as well claim “The brain is a finite state machine” because we can simulate them using pencil and paper.
Such modesty! It’s actually worse than that. You could write a program for a feedforward thermostat (i.e. which tries to predict how much to heat or cool based on factors other than the room temperature, like the sunshine, temp outside, insulation, etc.) on your computer, but Powers et al. would scream bloody murder if you tried to use that as evidence that living systems are feedforward control loops!
Actually, you couldn’t. At least, it wouldn’t work very well, not nearly as well as a system that simply measures the actual temperature and raises or lowers it as necessary.
Try it and see.
“Feedforward control loop” is pretty much a contradiction in terms. Look at anything described as feedforward control, and you’ll find that it’s wrapped inside a feedback loop, even if only a human operator keeping the feedforward system properly tuned. There are some demonstrable feedforward links in the nervous system, such as the vestibulo-ocular reflex, but as expected, the VOR is wrapped inside a feedback system that tunes its parameters. It wouldn’t work without that.
Ah, but if I deliberately created an artificial scenario designed to make FF control work, then FF control would look rockin’.
You know, like the programs you linked do, except that they pimp feedback instead ;-)
Yes, feedback control is usually better; my point was the excessive extrapolation from that program.
Yes, very true, which reminds me: I saw a point in the demo1 program (link when I get a chance) on the site pjeby linked where they have you try to control a system using either a) your knowledge of the disturbance (feedforward), or b) your knowledge of the error (feedback), and you inevitably do better with b).
Here’s the thing though: it noted that you can get really good at a) if you practice it and get a good feel for how the disturbance relates to how you should move the mouse. BUT it didn’t use this excellent opportunity to point out that even then, such improvement is itself due to another feedback loop! Specifically, one that takes past performace as the feedback, and desired performance as the reference.
PCT is not derived from the demos; the demos are derived from PCT.
So you see, wherever you look in the behaviour of living organisms, you find feedback control!
If that seems trivial to you, then it is probably because you are not an experimental psychologist, which is the area in most need of the insight that living organisms control their perception. You probably also do not work in AI, most of whose practitioners (of strong or weak AI) are using such things as reinforcement learning, planning, modelling, and so on. Robotics engineers—some of them—are about the only exception, and they have a better track record of making things that work.
BTW, I’m not touting PCT or anything else as the secret of real AI. Any better understanding of how real brains operate, whether it comes from PCT or anything else, will presumably facilitate making artificial ones, and I have used it as the basis of a fairly good (but only simulated) walking robot, but strong AI is not my mission.
In that case, why do you keep insinuating that control theory is useful as a high-level model of intelligence? That seems analogous to deriding computational neuroscientists for not using neurons in their models.
ETA: By comparison, evolutionary psychologists can’t do the math yet on the selective advantage of genes coding for variations in mental traits, but their derived models of human psychology allow them to make significant predictions that weren’t thought of without the model, and which often check out in experiments. Does PCT have any novel experimental consequences that have been verified?
I was expecting you to show me how the controls paradigm does a better job of explaining behaviors than the prevailing, but messy, model of evolutionary psychology. (Not that they would contradict, but your model would have to be simpler and/or more precise.)
That was the challenge I had presented to you before: show how “humans are control systems” is better at compressing a description of our observations.
If it seems hard to do, it’s probably because it is hard to, because it’s not a better model, because the behavior is so hard to express in controls terminology. Finding a mate is simply not like making sure one line is under another.
If I were the first person to publish the neuron theory, it would include an actual model with actual moving parts and therefore have actual explanatory power over actual observations that actual other people can make. It would not say, as you have essentially done, “people think with neurons, so like, when you think, it’s all … neurony. Are you thinking about sex? Yeah, the neuronal model has that. Neurons cause that thinking too. See, once I know what you’re thinking about, I can predict, using the neuronal model, what you’re thinking.”
But human behavior is where you’ve started, hence my confusion of what exactly the “humans as controllers” model accomplishes.
But it’s not an insight unless it makes the problem easier. Everything you’ve presented here simply shows how you could rephrase the solution to predicting organism behavior into a controls model once it’s been solved. So where does the ease come in? What problem becomes easier when I approach it your way?
I dismissed it as a trivial game because it is a trivial game. From the fact that I use proportional feedback control to keep two lines together, it does not follow that this is a useful general model of all organism activity.
Believe it or not, I have put a lot of work (in terms of fraction of my spare time) into exploring PCT. I’ve completed demo1 on the site pjeby linked, and have run it and four others in DOS Box. I’ve read some of the pdfs expalining feedback systems at the cellular level. I am going far out of my way to see if there’s anything to PCT, so please do not write me off as if I’m making you hold my hand.
I am making every effort to make things easy for you. All that’s left for your is to have a real model that you actually understand.
That, in turn, would rebut my strongly justified suspicion that the model’s “predictions” are ad hoc and the parallels with control systems superficial.
The reason why expressing the connection between not having a mate and seeking a mate in terms of PCT is so difficult is because “not having a mate” is not a perception, and because “seeking a mate” is not a behavior. Rather, these are an abstract world state with multiple perceptual correlates, and a broad class of complex behaviors that no known model explains fully. Given such a confusing problem statement, what did you expect if not a confused response?
The second problem, I think, is that you may have gotten a somewhat confused idea of what (non-perceptual) control systems look like. There was a series of articles about them on LW, but unfortunately, it stopped just short of the key insight, which is the PID controller model. A PID controller looks at not just the current value of its sensor (position, P), but also its recent history (integral, I) and rate of change (derivative, D).
If you want to test PCT, you need to step back and look at something simpler. The most obvious example is motor control. Most basic motor control tasks, like balancing, are a matter of generating some representation of body and object position, figuring out which neurons trigger muscles to push it in certain ways, and holding position constant; and to do that, any organism, whether it’s a human or a simple invertebrate, needs some neural mechanism that acts very much like a PID controller. That establishes that controllers are handled in neurology somehow, but not their scope. There’s another example, however, which shows that it’s considerably broader than just motor control.
Humans and animals have various neurons which respond to aspects of their biochemistry, such as concentrations of certain nutrients and proteins in the blood. If these start changing suddenly, we feel sick and the body takes appropriate action. But the interesting thing is that small displacements which indicate dietary deficiencies somehow trigger cravings for foods with the appropriate nutrient. The only plausible mechanism I can think of for this is that the brain remembers the effect that foods had, and looks for foods which displaced sensors in the direction opposite the current displacement. The alternative would be a separate chemical pathway for monitoring each and every nutrient, which would break every time the organism became dependent on a new nutrient or lost access to an old one.
Moving up to higher levels of consciousness, things get significantly more muddled. Psychology and clear explanations have always been mutually exclusive, and no single mechanism can possibly cover everything, but then it doesn’t need to, since the brain has many obviously-different specialized structures within it, each of which presumably requires its own theory. But I think control theory does a good job explaining a broad enough range of psychological phenomena that it should be kept in mind when approaching new phenomena.
I disagree, but that’s probably because I’ve seized on PCT as a compressed version of things that were already in my models, as disconnected observations. (Like time-delayed “giving up” or “symptom substitution”.) I don’t really see many gaps in PCT because those gaps are already filled (for me at least), by Ainslie’s “conditioned appetites” and Hawkins’ HTM model.
Ainslie’s “interests” model is a very strong fit with PCT, as are the hierarchy, sequence, memory, and imagination aspects of HTM. Interests/appetites and HTM look just like more fleshed-out versions of what PCT says about those things.
Is it a complete model of intelligence and humans? Heck no. Does it go a long way towards reverse-engineering and mapping the probable implementation of huge chunks of our behavior? You bet.
What’s still mostly missing, IMO, after you put Ainslie, PCT, and HTM together, is dealing with “System 2” thinking in humans: i.e. dealing with logic, reasoning, complex verbalizations, and some other things like that. From my POV, though, these are the least interesting parts of modeling a human, because these are the parts that generally have the least actual impact on their behavior. ;-)
So, there is little indication as to whether System 2 thinking can be modeled as a controller hierarchy in itself, but it’s also pretty plain that it is subject to the System 1 control hierarchy, that lets us know (for example) whether it’s time for us to speak, how loud we’re speaking, what it would be polite to say, whether someone is attacking our point of view, etc. etc.
It’s also likely that the reason we intuitively see the world in terms of actions and events rather than controlled variables is simply because it’s easier to model discrete sequences in a control hierarchy, than it is to directly model a control hierarchy in another control hierarchy! Discrete symbolic processing on invariants lets us reuse the controllers representing “events”, without having to devote duplicated circuitry to model other creatures’ controller hierarchies. (The HTM model has a better detailed explanation of this symbolic/pattern/sequence processing, IMO, than PCT, even though in the broad strokes, they’re basically the same.)
(And although you could argue that the fact we use symbols means they’re more “compressed” than control networks, it’s important to note that this is a deliberately lossy compression; discrete modeling of continuous actions makes thinking simpler, but increases prediction errors.)
What a pitiful excuse.
Let’s get some perspective here: the model I’m trying to understand is so vague in the first place (in terms of what insight it has to offer), despite all of my efforts to understand it with basic questions about what it does to replace existing models. Of course my questions about such an ill-supported paradigm are going to look confused, but then again, it’s not my responsibility to make the paradigm clear. That burden, currently unmet, lies on the person presenting it.
If you’re familiar with the tools of rationality, it is a trivial task to handle “confused” questions of exactly the kind I just asked—but that pre-supposes you have a clue what you’re talking about. All you have to do is identify the error it makes, find the nearest meaningful problem, and show how your model handles that.
A confused response is neither appropriate, nor deserved, and only reflects poorly on the responder.
Let me show you how it works. Let’s say I’m some noble defender of the novel Galilean model, trying to enlighten the stubborn Ptolemaic system supporters. Then, some idiot comes along and asks me, “Okay, smarty, how does the Galilean model plot the epicycle for Jupiter?”
In response, do I roll my eyes at how he didn’t use my model’s terminology and hasn’t yet appreciated my models beauty? Do I resign myself to giving a “confused response”? No.
And do you know why I don’t? Because I have an actual scientific model, that I actually understand.
So in response to such a “hopeless” dilemma, I marshal my rationalist skills and give a non-confused response.
Ready to have your mind blown? Here goes:
“When you ask me about Jupiter’s epicycle, what you’re really looking for is how to plot its position relative to earth. But my point is, you don’t need to take this step of calculating or looking up epicycles. Rather, just model Jupiter as going around the sun in this well-defined eilliptical path, and the earth in this other one. We know where they will be relative to the sun as a function of time, so finding Jupiter relative to the earth is just matter of adding the earth-to-sun vector to the sun-to-jupiter vector.”
There, that wasn’t so hard, was it? But, I had it easy in that I’m defending an actual model that actually compresses actual observations. Richard, OTOH, isn’t so lucky.
Notice what I did not say: “You find Jupiter in the sky and then you draw an epicycle consistent with its position, but with the earth going around the sun”, which is about what I got from Richard.
Yeah, that’s the point. Those higher levels are exactly what pjeby attempts to use PCT for, which is where I think any usefulness (of the kind seen in biochemical feedback loops) loses its compression abilities, and any apparent similarity to simple feedback control systems is superficial and ad-hoc, which is exactly why no one seems to be able to even describe the form of the relationship between the higher and lower levels in a way that gives insight. That is, break down “finding a mate” into related controllable values and identify related outputs. Some specification is certainly possible here, no?
Neither the problem statement, nor any of the confused responses were mine. My post was meant to clarify, not to excuse anything.
No, that is not the correct way to handle confused questions. The correct way to handle them is to back up, and explain the issues that lead to the confusion. In this case, there are many different directions the question could have been rounded in, each of which would take a fairly lengthy amount of text to handle, and people aren’t willing to do that when you could just say that wasn’t what you meant. I should also observe that pjeby gave you a citation and ducked out of the conversation, specifically citing length as the problem.
At some point, you seem to have switched from conducting a discussion to conducting a battle. Most of the parent post is not talking about the supposed topic of discussion, but about the people who participated in it before. Unfortunately, the history of this thread is far too long for me to read through, so I cannot respond to those parts. However, I am strongly tempted to disregard your arguments solely on the basis of your tone; it leads me to believe that you’re in an affective death spiral.
I know. Still a pitiful excuse, and yes, it was an excuse; you insinuated that my confused question deserved the flippant response. It didn’t. It required a simple, clear answer, which of course can only be given when the other party actually has a model he understands.
We’re bickering over semantics. The point is, there are more helpful answers, which one can reasonably be expected to give, than the “confused reply” you referred to. Richard knows what “finding a mate” means. So, if he actually understands his own model, he can break down “finding a mate” into its constituent references and outputs.
Or say how finding a mate should really be viewed as a set of other, specific references being tracked.
Or somehow give a hint that he understands his own model and can apply it to standard problems.
Was my epicycle example not the kind of response I could reasonably expect from someone who understands his own model?
That’s because, as I said, Richard completely handwaved the PCT explanation, in the same way as an evolution supporter would likely end up handwaving ideas like “inclusive genetic fitness”.
To model mate searching in PCT, you would need to include several continuous variables related to mate selection, such as “how attractive am I” and “how often am I getting laid”, that would be used to control other variables like for “minimum required desirability-of-partner”. Lumping that all together into one variable as Richard suggested would be quite ludicrous. (But note, by the way, that there’s mainstream evidence for the existence of such variables in humans and other animals.)
Keep in mind that controllers are hierarchical, so low-level behavior patterns are driven by higher level ones. So, the “minimum partner desirability” threshold would gate the program-level controller for mate attraction behaviors… which would include various levels to track of the potential mate’s apparent response or interest, etc. You’ve got levels going all the way down here, implied by “hierarchical control”, in the same way that a huge host of behaviors and characteristcs are covered by “inclusive genetic fitness”.
I did that, in precisely the comment you linked to as saying I didn’t. Specifically, I pointed out that the exact set of definitions for e.g. “thinking hard” that wouldn’t work in reality, precisely match the ones that PCT predicts would not work, and vice versa for the definitions that would not work.
If you’re saying, you don’t see how that’s so from what I wrote and need more detail, that’s fine. But to me, I did exactly as you requested, or the best I could given your undefined phrase “thinking hard”.
You quoted me out of context; the rest of the comment goes on to explain that what I find useful is that PCT tells us what things to look for in order to solve the problem.
What I perhaps didn’t explain very well is how those “things to look for” differ from what might otherwise be looked for. For example, PCT emphasizes integrated continuous variables rather than discrete events, e.g. “probability of receiving a shock within time period t”. Our normal thinking about behavior emphasizes discrete goals and actions, while one of PCT’s central ideas is that discrete goals and actions occur only to restore continuous variables to their reference ranges, in response to either environmental disruptions, or the passage of time (causing an integrated or “average” level to drop).
No, not like an evolution supporter, because an evolution supporter could identify what exactly IGF refers to in a way that is not a trivial restatement of the problem of “Why would a mother give up her life for her two children?”
A scientific-sounding answer would be, “because that improves her IGF”, and you’re correct this would be a handwave.
A really scientific answer would “because that improves her IGF, which is roughly the fraction of the next generation’s genes that are hers, which would account for why the gene pool is dominated by the genes of people who made such a tradeoff”.
Richard does not have such a moving-parts model that breaks down the concepts of “having a mate” and “recognition of having a mate” and “action that moves me toward having a mate”. They’re irreducible black boxes, at least as far as he’s aware. That’s not a theory. If he had a theory, he should have and would have focused more on expanding one of these concepts than on making sure the problem was fully restated with with different terms.
in other words, give an answer different from RIchard. Hint: if the very same model can be “interpreted” to imply different explanations, it’s not a good model nor a theory.
No, you asserted it, with no reference to the specific aspects of PCT that rule against the “non-phenomena”.
Surely my query warranted more of an answer than, “yep, everything’s okay over here” even if you went through the effort to transform it to, “yep, what I’ve seen work, it sure is expected to work by PCT, and what I know doesn’t work, yep, it sure is claimed by PCT not to work!” (Edit: removed some snarkiness)
The context makes you look worse! Here it is:
So basically, “all” you have to do is find the reference values and appropriately modify them … but specifying the reference values itself contains the exact same insights you’d need to solve the problem anyway! (it’s “AI-complete” in compsci/AI jargon) So the “human as controller” model doesn’t simplify the problem, it just says “here, go solve the problem, somehow, and when you do, without the help of this model, you’ll see that one of the six trillion neat things you can do is specify them in controls format”.
The fact that you’re still ignoring any of the substantive and responsive portions of my comments, bodes ill for this being a useful exchange.
It is quite possible I’ve misunderstood your queries and/or answered them inadequately. However, I’d like to think that the appropriate response in that case would be to clarify what you want, rather than simply taking it to mean no-one can give you what you want.
I’ve mentioned a number of things that PCT does beyond that. For example, it shows that the first thing to look for in modeling are continuous analog variables integrated over a time period, with shorter time periods generally being represented lower in the control hierarchy than longer time periods. AFAICT, this is far from an obvious or trivial modeling distinction.
The fact that you chose not to comment on that, but instead dug in on justifying your initial position, suggests to me that you aren’t actually interested in the merits (or lack thereof) of PCT as a modeling tool, so much as in defending your position.
Yeah, I like that strategy. “In this extremely long, involved exchange, any part of my post that you didn’t directly respond to, was an ultra-critical omission, and completely demonstrates your failure to act in good faith or adequately respond to my points.”
Whatever. I didn’t respond to it because I haven’t gotten around to responding to your points in the other thread, or because it didn’t address my request. In this case, it’s the latter.
Okay, and what epistemic profit does this approach gain for you, especially given that deliberate actions in pursuit of a goal are highly discontinuous? Oh, right, add another epicycle. Hey, the Hawkins HTM model, that’ll work!
ETA: Do not interpret this post to mean I’m in full anti-PCT mode. I am still exploring the software on the site pjeby linked and working through the downloadable pdfs. I’m making every effort to give PCT a fair shake.
I’m not certain I understand your terms. If I interpret your words “classically”, then of course I “know what you mean”. However, if I’m viewing them through the PCT lens, those words make no sense at all, or are blatantly false.
When you drive a car and step on the brake, is that a “deliberate action” that’s “discontinuous”? Classically, it seems obvious. PCT-wise, you’re begging the question.
From the PCT perspective, the so-called “action” of braking is a chain of controls looking something like:
Speed controller detects too-high speed, sets speed-change controller to “rapid decrease”
Speed-change controller detects discrepancy between current acceleration and desired deceleration, sets braking controller to “braking hard”
Braking controller notes we aren’t braking, sets foot position to “on brake”
Foot position controller detects foot is out of position, requests new leg position
Leg position controller detects out of position, requests new leg speed/direction
Leg speed controller detects not moving, requests increased muscle force
...etc., until
Foot position controller detects approaching correct position, and lowers requested movement speed, until desired position is reached
Speed controller observes drop of speed below its reference level, sets speed-change controller to “slow accelerate”
Speed-change controller notices that current deceleration is below “slow accelerate” reference, sets “gas” controller to “slight acceleration”
...and so on, until speed stabilizes… and the foot goes up and down slightly on the gas… all very continuously.
So, there is nothing at all “discontinuous” about this. (Modulo the part where nerves effectively use pulse-width modulation to communicate “analog” values).
And it’s precisely this stable continuity of design that makes PCT so elegant; it requires very little coordination (except hierarchically), and it scales beautifully, in the sense that mostly-identical control units can be used. Got a more complex animal? Need more sophisticated behavior? Just add controllers, or new layers of controllers.
Need a new skill? Learning grows in or assigns some new controllers, that measure derived perceptual quantities like “speed of the car”, “braking”, and “putting on the gas”. (Which explains why procedural knowledge is more persistent than propositional knowledge—the controllers represent a hardware investment in knowledge.)
And within this model, actions are merely side-effects of disturbances to the regulated levels of perceptual variables, such as speed. I stopped the upward point of the hierarchy at the speed controller noticing a speed discrepancy, but the reason for that discrepancy could be you noticing you’re late, or it could be that your “distance to next car” controller has issued a request to set the new “desired speed” to “less than the car in front of us”. In either case, the “action” is the same, regardless of what “goal”—or more likely, disturbance—caused it to occur.
That being said, PCT does include “sequence” and “program” controller layers, that can handle doing things in a particular sequence or branching. However, even these are modeled in terms of a perceptual control hierarchy, ala TOTE loops. That is, you can build TOTE loops by wiring controllers together in relatively simple ways.
Reification of programs and “actions” through controller hierararchies is also a good strategy for building a fast machine out of slow components. Rather than share a few ultra-fast, complex components, PCT hierarchies depend on chains of similar, simultaneously-responding, cheap/dumb components, such that the fastest responses are required from the components that are generally nearest (network-wise) to the place where the signals need to be received or delivered to exert control.
These are just some of the obvious properties that make PCT-style design a good set of tradeoffs for designing living creatures, using similar constraints to evolution. (Such as the need to be able to start with primitive versions of the model, and gradually scale up from there.)
As I said, it gives me a better idea of what to look for. After grasping PCT, I was able to identify certain “bugs” in my brain that had previously been more elusive. The time and hierarchy distinctions made it possible for me to identify what I was controlling for, rather than just looking at discrete action triggers, as I did in the past.
In this area, PCT provides a more compact model of what psychologists call “secondary gain” , hypnosis people call “symptom conversion”, and NLP people call “ecology”.
The idea is that when you take away one path for someone to get something (e.g. giving up smoking) they may end up doing something else to satisfy a need that was previously supported by the old behavior (e.g. chewing gum).
What psychologists, NLPers, and hypnosis people never had a good explanation for (AFAIK) is why it takes time for this substitution or reversion to occur! Similarly, why does it take time for people to stop persisting at trying to do something new?
This is an example of a complex behavioral property of humans that falls directly out of the PCT model without any specific attempt to generate it. Since high-level goals are integrated over a longer time period, it takes time for the error signal to rise, and then further time for the controller network reorganization process (part of the PCT model of learning) to find an alternative or extinguish the changed behavior.
I find PCT parsimonious because there are so many little quirks of human nature I know about, that would be naturally expected to occur if behavior was control-system driven in precisely the ways PCT predicts that it is… but which are just weird and/or unexplained under any other model that I know of.
Okay, thank you, that was exactly the kind of answer I was looking for, in terms of breaking down (what is framed by us non-PCTers as) a discrete list of actions into hierarchical feedback loops and what they’re using for comparison. Much appreciated.
But just the same, I think your explanation illuminates my complaint about the usefulness of the model. What it appears to me is, you just took a list of discrete steps and rephrased them as continuous values. So far, so good, but all I see is added complexity. Let me explain.
I would describe my steps in baking a cake (and of course this abstracts away from lower level detail) as:
1) Open preheated oven.
2) Place pan containing batter onto middle of middle over rack.
3) Close oven.
4) Set timer.
Your claimed improvement over this framing of these events is:
1) Define variable for oven openness. Recognize it’s zero and push it toward 1.
2) Define variable for pan distance from middle of middle oven rack. Recognize it’s too high and push it toward zero.
3) Recognize oven openness is 1 and should be zero, push it in that direction.
4) Define variable for oven-timer-value-appropriateness. Recognize it’s too low and move it higher.
Yes, superficially, you’ve made it all continuous, but only by positing new features of the model, like some neural mechanism isomorphic to “detection of oven-timer-value-appropriateness”, which requires you to expand that out into another complex mechanism.
I agree, as I’ve said before, that this is one way to rephrase what is going on. But it doesn’t simplify the problem; it forces you identify the physical correlate of “making sure the oven’s set to the right time” in a form that I’m not convinced is appropriate for the problem. Why isn’t it appropriate?
Among other things, you’re forced to solve the object recognition problem and identify a format for comparison. But if I’ve solved the (biological) object recognition problem, my model can simply invoke the actual neural mechanism being used, without the added complexity of reformatting the causal flow into feedback loops.
You defend this model by its elegance, but you only get the elegance after you solve the problem some other way. That is, I only have an elegant hierarchical feedback loop if I can, somehow, solve the object recognition problem that allows me to actually specify a reference and feedback signal. A model isn’t any good if it presupposes the solution of the problem it’s being used to solve.
Hope that clarifies where I’m coming from.
You could describe them that way, yes, and that would nominally describe the behaviors you emit. However, it’s trivial to prove that this does not describe the implementation in your head that emits those behaviors!
For example, you might forget to preheat the oven, in which case the order of your steps is going to change. There are any number of disruptions that can occur in your sequence of “steps”, that will cause you to change your actions to work around the disruptions, with varying degrees of automaticity, depending on how high in your control hierarchy the disruptions reach.
A simple disruption like a spill on the floor you need to walk around will be handled with barely a conscious notice, while a complex disruption like the power being off (and the oven therefore not working) will induce more complex behavior requiring conscious attention.
If a sequence of steps could actually describe human behavior, we could feed your list of steps to a computer and get it done. It’s actually omitting some of the most important information: the goals of the steps, and how to measure whether they’ve been obtained.
And that’s information our brains have to be able to know and use in order to actually carry out behavior. We tend to assume we do this by “thinking”, but we usually only “think” in order to handle high-level disturbances that require rearrangement of goals, rather than just using existing control systems to work around the disturbance.
When you get to higher levels of modeling, you can certainly deal with sequences of subgoals that are matching of perceptual patterns like “cake is in the oven”. Did you read the TOTE loops reference I gave? TOTE loops act on something until it reaches a certain state, then activate another TOTE loop. PCT incorporates the previously-proposed cog psych notion of TOTE loops, and proposes some ways to build TOTE loops out of simpler controllers.
Part of the modeling elegance of using controllers and TOTE loops to implement overall behavioral programs is that they allow you to notice, for example, that your assistant chef has already set the timer for you as you placed the cake in the oven… and thereby skipping the need to perform that step.
This I don’t get. We already know that (visual) object recognition can be implemented by time-varying hierarchical feature sequences tied to the so-called “grandmother neurons”. Both HTM and PCT include this concept, except that PCT proposes you get an analog “grandmotherness” signal, whereas IIRC the HTM model assumes it’s a digital “grandmother present” signal. But HTM at least has pattern recognition demos that handle automatic learning and pattern extraction from noisy inputs, and it uses exactly the same sort of recognition hierarchy that the full PCT model calls for.
That’s why I keep saying that if you want to see the entire PCT model, you need to read the book. Most of the primers either talk about low-level stuff or high-level stuff. Object recognition, action sequences, and that sort of thing are all in the middle layers that make up the bulk of the book.
Note, by the way, that this is backwards, when applied to listing steps and calling it a model. The steps cannot be used to actually predict behavior, because they list only the nominal case, where everything goes according to plan. The extra information that PCT forces you to include results in a more accurate model—one that does not simply elide or handwave away the parts of behavior that we intuitively ignore.
That is, the parts we don’t usually bother communicating to other human beings, because we assume they’ll fill in the gaps.
PCT is useful because it shows where those gaps are and what is needed to fill them in, in much the same way that the initial description of evolution identified gaps in our knowledge about biology, and what information was needed to fill them in, in place of the asumptive handwaving that “God did it”. In the same way, we currently handwave most of what we don’t understand about behavior as, “X did it”, where X is some blurry entity or other such as learning, environment, intelligence, habit, genetics, conditioning, etc.
(And no, PCT doesn’t merely replace X with “control systems”, because it shows HOW control systems can “do it”, whereas other values of X are simply stopping the explanation at that point.)
see correction