Dick Robertson, catching up

[From Dick Robertson, 2010.09.22.0254CDT]
Hi Bill,

I finally found your 9/14/10 response to my attempt to get a discussion of what a model of a higher level would look like, and will study it carefully as I can see it gets into the heart of my interest. I also found your post about getting me on to some list. I think it has to do with your paper on the Paradigm, which I have just finished reading. I seem to be quite behind on this thread, as i was on a trip to visit relatives in MN and then to the Mayo Clinic, getting back on 9/14 and finding the PCT world had raced ahead of me. I did write that piece about HPCT modeling in a motel one night. But now that I’ve read the paradigm – and liked it a lot, although I could make a few editorial comments, if they are apropos. I also find a discussion of a tracking study that Martin and Rick are examining, but I can’t find the original task, or is it taking from an earlier report?

Anyway, I’m trying to catch up. What should I do to justify being placed on the list you first talked about.

Hopes this makes some sense.


Dick R


----- Original Message -----
From: Bill Powers powers_w@FRONTIER.NET
Date: Tuesday, September 14, 2010 9:58 am
Subject: Re: Perceptual Controlled Variables

[From Bill Powers (2010.09.14.0655 MDT)]

Dick Robertson,2010.09.14.0845CDT –

DR: I think Gavin Ritz has a point when he raises his question about HPCT. That is his claim (if I understand correctly) about the difficulty of describing a proper model for higher levels of controlled variables.

BP: Gavin is, of course, right about the difficulties with the higher levels in the HPCT model. But he has misidentified the problem. There are really two main problems: first to define what the higher levels do, and second to explain how they do it. I have focused almost entirely on the question of what the higher levels do, with only occasional glimpses of answers to the other question: how they have to be organized to do it. But this doesn’t make my approach different from anyone else’s: nobody knows even how configuration perception works, much less any higher level. The nearest anyone has come to artificial perception of configurations is to develop systems that can classify and name configurations: this is called “apple,” that is called “zebra.” But that doesn’t explain how a person can recognize and control a cube balanced on one corner and spinning, which doesn’t require classifications or names.

We don’t need to know how perceptual input functions work to experience their outputs. I’m glad to see that you have arrived at the same conclusions.
We all know what a configuration is, and I think there’s little argument about the general definition; something made of different sensations at Level Two, but which is not itself a sensation.

This is true of all the other levels, as far as I’ve been able to find consistent and mutually-coherent definitions. I don’t think anyone can argue that we don’t perceive sequentiality (ordering), or relationships, or logical functions (one if by land, two if by sea), and so on. I haven’t yet heard from anyone who simply doesn’t perceive the types of perceptions I’ve defined. Well, maybe one or two have expressed puzzlement about what I mean by “principles.”

That doesn’t mean I’m sure I’m right, and I never encourage anyone to memorize the levels as I have defined them, or to try to find those levels in applications like the Method of Levels. The idea of relative level is enough for now.

DR: That and the other recent thread highlighted the need for a definition of what a model is. My idea derives from the very literal models we looked at in grade school, the model of the solar system in which you could see the relationships between the sun and planets. You could see why – as long as you thought the earth the center – the planets would sometimes appear to move backwards. You could see the phases of the moon as it moved around the world (our model had a light bulb for the sun in the middle. There were other things that helped define the concept of model for me, model trains, for example. What all these “models” had in common was that they were physical objects that corresponded in essential ways to the larger objects – systems – to which they applied. I don’t recalll seeing Gavin’s definition of a true model, but I assumed it might be something like that too. So when he decried the lack of analogues to that concept in HPCT I had to agree.

Perhaps a better word than model would be simulation. A simulation is a working representation of the relationships among variables in a physical system. To simulate Newton’s Laws of Motion, for example, we would start with the famous F = MA, which states that Acceleration is directly proportional to applied Force, the constant of proportionality being called Mass. A signal representing force/mass can be fed into an integrator to produce an output corresponding to velocity; a velocity signal can be fed into a second integrator to produce an output corresponding to position. So without trying to include irrelevant variables like price, odor, color, size, or shape, we can simulate the behavior of a free-floating object when it’s subjected to an applied force. The simulation will obey Newton’s laws of motion, and from its behavior we can predict the behavior of a real physical system. That’s the object of modeling: to construct a simulation which behaves in the relevant ways as much like the real system as possible. By seeing how the simulation has to be organized to achieve this end, we learn how the real system is organized.

So modeling is not just minaturizing. If size made a difference in the behavior we would have to give the model the correct size, too (for example, if we want to take air resistance into account). But we don’t have to construct an actual physical duplicate of the system. We can do all this with computations: we can use numbers for F, M, and A, and mathematical functions such as F = MA to represent the relationships among these variables that are actually observed, measured. We don’t have to make the computation weigh 10 kilograms to compute the acceleration of a 10-Kg mass by a force of 10 Newtons.

DR: When we get above the level of relationships, the controlled variable of interest is wholly in one’s head. I believe. We do not automatically see the same categories as the next person. We can train each other to a point of considerable agreement.

BP: This is partially true, partially false. Consider how you manage to enter a house through the front door. You reach out, grasp the knob, pull the door open, and walk through. While this sequence may exist only in your head, you will find that no other sequence will work; you can’t, for example, walk through the door and then open it. While you can’t put your finger on any actual material thing that corresponds to the sequence, it nevertheless has an existence independent of you and you have to control the perception of sequence in the right way if you want to enter the house.

All of the levels I have defined have this nature: they are perceptions that must be controlled in particular ways in order to have the effects that are wanted. We have little choice in this: we learn the rules or we fail to control. So there must be something objective about whatever underlies these levels. That is what gives the definitions some degree of reality. These levels are clues about our connections to whatever it is that lies beyond the senses.

DR: So are we still without models for higher orders?

BP: We are not without models of the behavior of higher orders of control. You and David Goldstein used a model of self-concept in your study – you used the principles of negative feedback control to predict what a person would do when self-descriptions were contradicted, and found that the predictions were nearly 100% correct.

What you didn’t have was a model of how a self-concept perception is derived from lower-order perceptions. For that, you had to rely on your own ability to perceive self-concepts in relationship to statements about personal characteristics. You could perceive the lower-order elements and you could perceive the effects on a higher-order perception, so even though you couldn’t fill in the mechanism by which the higher-order perception was generated, you could still predict from your own experiences what the perception would be. In the same way, I can’t tell you how a spinning cube is perceived in the midst of a huge collection of changing sensations, but I can at least test to see if another person perceives and controls the same thing under the same conditions in which I perceive and control it. This shortcut seems, in fact, to work quite well. You’ll notice that in LCS3, I predict quite confidently exactly how a reader will behave in all the interactive demonstrations included in the book – and I would be surprised if those predictions were ever wrong.

DR: If these are models of higher order systems, then, even if we haven’t yet found out how to construct them on computers it is obvious that that will come. After all robots are coming closer to simulating human behavior in many respects. When they do so aren’t they being models?

BP: It’s not necessary to use a computer or build a robot to construct and test models. The only real requirement is that the behavior of the model follow strictly from the properties you give it, so you can’t reach in and tweak it if it starts to misbehave. The main advantage of a computer is that you can program it in advance so it can ONLY do what you have specified that it does; this means you can’t cheat. But if you don’t want to cheat, and really want an honest test of a model, all you have to do is specify very carefully the properties you are giving the model, using unambiguous language set down on paper so it can’t be changed after you start to apply the model. You have to specify how to reason about the model, how to make predictions. Then all you or anyone else needs to do is apply the rules you have stated, without any deviations at all, even when you can see that the predicted behavior is departing from the real behavior.

That is really the hard part. Making a committment to a model sounds easy until you start to have hopes that it will work. No model ever works exactly as you thought it would the first time. You need the self-discipline to avoid changing your premises or rules as you see failure looming up. You have to avoid taking the easy way out by using weasel-words whose meanings can be changed as needed without changing the words, so it sounds as if you’re still making the same prediction as before, only this time it’s right.

I’m pretty good at cheating like that, which is why I stick to computer models. Deep down, I would really rather know the truth than be right. Also, when you construct a computer model that you can’t possibly interfere with once it’s running, there is no feeling in the world like seeing the model do exactly what you thought it would do down to the last significant digit. A few of us have experienced that; it doesn’t matter that the models were simple, it’s still a miracle when you make contact with actual reality in this way. It’s a rare experience and worth all the effort to get there.


Bill P.