A skeptical engineer joins CSGnet. Albus.
Unedited posts from archives of CSG-L (see INTROCSG.NET):
A SKEPTICAL ENGINEER CHECKS OUT PCT AND CSGnet.
This is a rather complete thread of Paul George discussing Albus withCSGnetters and appreciating _Behavior: The Control of Perception_ towards theend. As always, every discussion provokes thoughtful posts. Note TomBourbon's post: Albus articles and Bill Powers' post: PCT models andsimulations.
I hope that this thread will prove helpful as an introduction to PCT andCSGnet.
March 21 1995, Dag
------------------------------------------------------------
Date: Thu Jun 23, 1994 4:39 pm PST
Subject: PCT models
I have been lurking for a while listening to the PCT debates, and think youmay be using too simplistic a view of a control system, which is biasing yourdiscussion (in fairness I haven't had access to the books/papers on thesubject, just posts on BPR_L by Dag).
Sophisticated control systems don't use reference variables (e.g. setpointsor alarm thresholds) they use reference _models_ (reflected in control logic).In a sense we have a continuously running simulation of the 'real' world towhich we compare our perceptions (sensor inputs). As PCT stresses we alsofilter our perceptions based upon our current model, trying to separate signalfrom noise.
The basic components of this view of a control system are SensoryProcessing (filter/transducer), Value Judgement (comparator), WorldModel/database, and Behavior Generator. Part of the model is some set of goalsor desirable states we wish to maintain or approach. We modify the model tobetter reflect 'outside reality' (as perceived) so that we can better predictthe results of our actions (actuator outputs).
For a good exposition of this model I suggest:
"Outline for a Theory of Intelligence" James S. Albus IEEE Transactions onSystems, Man, and Cybernetics, vol 21 #3, May/June 1991, p 473-509. IEEE log #9042583
Date: Fri Jun 24, 1994 7:28 am PST
Subject: Re: PCT Models
[Paul George 940624 0900] >[Martin Taylor 940623 20:30]
> Sometimes these more complex control systems do get talked about, butone of the questions of interest is what can be done with the simplestversions. For example, if we assert that the simple elementary control systemcontrols a scalar variable, it is possible to imagine that variable being theoutput of a neuron. Is there any evidence to suggest a need for more complexcontrolled variables in real systems? Is it not always possible to emulate theeffect of a complex control system with a hierarchy of scalar controlsystems?
{snip}
Basic general systems theory and gestalt psychology indicate that thesimple reductionist approach doesn't work well for complex systems,particularly biological ones. The whole is more than the sum of the parts.There is a qualitative difference between a Cray and a flip/flop (bistablemultivibrator for the picky).
I may be covering old ground, but simple scalar variables are notsufficient for control of dynamic processes, except in a simplistic case. Weare here dealing with human behavior. Even for simple astrodynamics or objecttracking you need vector systems such as Kalman Filters. A model need reflectpertinent aspects of the real world."Keep it is simple as possible, but nosimpler".
Actually Albus posits a hierarchy of simple control systems where eachcontrol component interacts with like nodes both up and down the hierarchy.Effectively nodes form a network, incidently mapping to the structure ofnervous systems.
> The basic attitude of a scientist when confronted with an explanationmore complex than the one in current use is "show me what it explains beyondwhat I already explain, or more precisely than I already explain it."
Albus's model adds several things that allow separation of concerns and anacknowledgement of internal complexity. It takes us beyond what might beunfairly called a black box model.He alleges that it has been validated interms of actual biological and behavioral structures (I haven't backtracked hisreferences, but matches my understanding of such mechanisms).
1) a specific locus for a atonomic perceptual filter & sampler. Thesimplistic diagram seems to take a 'then a miracle occurs' approach. This helpsdeal with the fact that we don't mentally control all perception. Some thingsare wired in. Second we do have some control on sensitivity levels and whichsensory inputs we pay attention to at a particular time. This has beenexperimentally verified.
2) An explicit concept of both reference variables and some kind of a statemachine (or other rule/history mechanism) to allow prediction and complexbehavior. "World model" overstates the concept somewhat, especially at thelowest level of the hierarchy. It is a history/prediction mechanism adequate tothe task at hand. It also helps provide a locus for simple fight/flightresponse mechanisms and 'habitual' responses. The Model node can interactdirectly with sensory perception and behavior generation nodes with out themediation of Value Judgement.
3) A specific locus for goal seeking behavior and evaluation internal modelvs external inputs and planned outputs. This is where either simple thresholdevaluation or complex rule processing takes place. It is, if you will the seatof 'consciousness' or 'will'.
4) An explicit locus for patterns of action or outputs. This cover bothinstinctual or otherwise hard wired actions (pulling a hand back from a flame),learned patterns (riding a bike), and consciously defined action plans. It alsoprovides a mechanism for Lilly's behavioral metaprograms.
{World model component description}
> I accept that such modification is required, and have several timesintroduced it explicitly into diagrams. But I do not know that it cannot beequivalently represented in the interconnections of a hierarchy of controlsystems without explicit models. To introduce an explicit model is tointroduce the question of mechanism of data storage and data retrieval.
I am not sure you need to specify a mechanism, though it is an interestingquestion, just recognize that it appears to exist. We acknowledge that wepossess memory and can use language, even if we don't understand how it worksat the neural level. Distributing the memory through the control systemhierarchy just means a distributed model which incorporates the control logic.This is not wrong, but appears to hide the issue. I aesthetically prefer theconcept of a hierarchy of models and value judgement systems. Whether you usethis model or a black box +ghost in the machine' approach depends on the issueyou are discussing. From the standpoint of observing external behavior, itdoesn't matter. But the internal structuring may help our understanding of whatis going on inside, from a conceptual standpoint.
Paul C. George
Date: Wed Jun 29, 1994 5:17 pm PST
Subject: Albus and Powers
From Tom Bourbon [940629.1731]
I'm just now working my way through the thread on "PCT models" initiated byPaul George (23 June 1994). Paul cited a 1991 article by Albus, and I noticethat Martin Taylor has copied the article to read. I think a good place tocompare Albus and Powers is in their independently written series that ran in_BYTE_ magazine, in June, July, August and September, 1979. Bill Powers wroteabout the CST model (as PCT was known back then) and told how to actually writeprograms to test the model. Bill's descriptions of the model and the programswere crystal clear. His programs, hence his model, actually _worked_ -- a featthat some of you know is of great significance to the modelers on thisnet.
In contrast, Albus began the fourth article in his series (which wastitled, "A model of the brain for robot control") by saying, "The essence of ahierarchy is that control is top-down. The ultimate choices are made at thetop, and the goals selected at this level are decomposed into action as theyfilter down through the various levels of the hierarchy." That's about as faras you can get from the idea about a control hierarchy in PCT; top-downhierarchies of that kind don't work as models for living systems. The remainderof the Albus article contained no working model for a robot, or for a brain. Itdid contain lots of speculation and guesses about the future of robots.
I recommend the two parallel, but widely diverse, series of articles toanyone who wants to compare the ideas of Powers and Albus.
Later, Tom
Date: Wed Jun 29, 1994 5:24 pm PST
Subject: Re: PCT models
From Tom Bourbon [940629.1757]
I've finished reading the new thread on PCT models. Here's a first smallset of comments.
In Message Thu, 23 Jun 1994 12:04:28 EDT, Paul George writes:
> I have been lurking for a while listening to the PCT debates, andthink you may be using too simplistic a view of a control system, which isbiasing your discussion (in fairness I haven't had access to the books/paperson the subject, just posts on BPR_L by Dag).
OK. You haven't read any of our work but you think it is too simplistic andit doesn't work. Since you are new to this net, we don't know about your work,either. Can you tell us something about how you decide whether a model works,or not? It would be helpful to me if you were to say some things about thatquestion in general, then tell us specifically how you apply those criteria toquantitative predictions made with the Albus model and why you think theresults lend credence to it.
> Sophisticated control systems don't use reference variables (e.g.setpoints or alarm thresholds) they use reference _models_ (reflected incontrol logic). In a sense we have a continuously running simulation of the'real' world to which we compare our perceptions (sensor inputs). As PCTstresses we also filter our perceptions based upon our current model, trying toseparate signal from noise.
Don't worry. We make no pretense of sophistication, so that field is wideopen to anyone who wants to claim it!
Later, Tom Bourbon
Date: Thu Jun 30, 1994 7:05 am PST
Subject: Re: PCT Models
From Tom Bourbon [940630.0823]
>[Paul George 940624 0900] >>[Martin Taylor 940623 20:30]
Martin:
>> Sometimes these more complex control systems do get talked about,but one of the questions of interest is what can be done with the simplestversions. For example, if we assert that the simple elementary control systemcontrols a scalar variable, it is possible to imagine that variable being theoutput of a neuron. Is there any evidence to suggest a need for more complexcontrolled variables in real systems? Is it not always possible to emulate theeffect of a complex control system with a hierarchy of scalar controlsystems?
Paul:
> Basic general systems theory and gestalt psychology indicate that thesimple reductionist approach doesn't work well for complex systems,particularly biological ones.
Could you tell us something about _how_ systems theory and gestaltpsychology "indicate" that a "reductionistic approach doesn't work for complexsystems" and _why_ you believe those ideas would apply to our work -- which youhave not read? I am not familiar with the process of "indication" and do notknow how to use it when I compare the behavior of various generative models.I'm in the dark about what you mean and will remain there until you say a fewthings about how you test the accuracy of quantitative predictions from theAlbus model.
> The whole is more than the sum of the parts.
I believe the original statement was closer to, "The whole is _different_from the sum of the parts." Whichever way we say it, so what? How can thattruism guide us in evaluating the performance of generative models of behavior?Precisely how does it help us compare, to pick an example out of the air, thetwo models described independently by Albus and Powers, in 1979?
> There is a qualitative difference between a Cray and a flip/flop(bistable multivibrator for the picky).
Yes. So?
> I may be covering old ground, but simple scalar variables are notsufficient for control of dynamic processes, except in a simplistic case. Weare here dealing with human behavior.
Yes, we are indeed dealing with human behavior here. When we use our simple(and insufficient) PCT scalar models to predict the performance of humans in arange of tasks, the predictions account for over 99% of the variance in thedata from the humans. How well does the (sufficient) Albus model do in suchcases? It must be pretty near perfect, if your remarks about it are close tothe facts.
> Even for simple astrodynamics or object tracking you need vectorsystems such as Kalman Filters. A model need reflect pertinent aspects of thereal world."Keep it is simple as possible, but no simpler".
_Who_ needs vector systems? We (PCT modelers) don't, at least not just yet.How much more of the variance does the Albus model "explain" when it isequipped with them than without them?
"A model need reflect pertinent aspects of the real world." Why? What isthe evidence for this statement, from testing models? What if there are modelsthat perform very well but do not "reflect pertinent aspects of the realworld." Which do we question or reject, the models, or the dictum?
> Actually Albus posits a hierarchy of simple control systems where eachcontrol component interacts with like nodes both up and down the hierarchy.Effectively nodes form a network, incidently mapping to the structure ofnervous systems.
Albus posits a top-down hierarchy. Have you tested a generative version ofa top-down model? In our experience, they don't work as models of livingsystems. We would like to see evidence to the contrary if it isavailable.
Paul:
> Albus's model adds several things that allow separation of concernsand an acknowledgement of internal complexity. It takes us beyond what might beunfairly called a black box model.
Martin Taylor asked you some good questions about those remarks.
> He alleges that it has been validated in terms of actual biologicaland behavioral structures (I haven't backtracked his references, but matches myunderstanding of such mechanisms).
I noticed that, at several points in your posts on this thread, you said(or implied) you hadn't actually checked out Albus's allegations and claims.Have you seen any simulations or quantitative predictions from him? Have yourun any of your own, using his model? I'm asking seriously -- you are not thefirst person to tell us that Albus says something important about the phenomenawe study in PCT. We are still looking for good evidence to support thoseassertions.
Your enumeration of the following items from the Albus model makes it clearthat the parallel articles in _Byte_, in 1979, are still one of the bestsources for a comparison of Albus and Powers. You are talking about a differentbreed of cat from the PCT model.
> 1) a specific locus for a atonomic perceptual filter & sampler..
> 2) An explicit concept of both reference variables and some kind of astate machine (or other rule/history mechanism) to allow prediction and complexbehavior. . . .
> 3) A specific locus for goal seeking behavior and evaluation internalmodel vs external inputs and planned outputs. . . .
> 4) An explicit locus for patterns of action or outputs. . .
No thanks. I'll stick with functional, generative PCT models. Call mebackward and inadequate; I don't care. ;-) (I _will_ rethink that position if Isee examples of generative models that adhere strictly to the Albusarchitecture _and_ produce fits to quantitative human data that are as good asor better than the fits from simple PCT models.)
Later, Tom
Date: Thu Jun 30, 1994 8:51 am PST
Subject: Re: PCT models
I didn't say that PCT models wouldn't work. I just suggested a more complexinternal organization might increase understanding, particularly in terms ofwhere communications errors may occur. I do not judge PCT, as I don't yet knowa lot about it. On the surface it seem very attractive. I laud the communitieswork in attempting to validate the theory via modeling.
A model is valid if it is useful for understanding and mimics the behavioror a portion of the real system. To simple a model can blind you tocomplexities of the real system and subtleties of its behavior. Perceptualfiltering tends to make you ignore things you weren't expecting or were outsidethe parameters of your mental model.
Modeling is a useful technique but subject to limits, particularly inselecting the 'important' characteristics of the real system. Observing livingsystems and man-made control systems and generalizing to produce theory is alsovalid. Validation by inspection is a valid technique. Albus didn't make'quantitative predictions', he proposed structure and semantics.
On the whole you appear very defensive, overreacting to perceivedcriticism. I would hope this group is not limited to TRUE BELIEVERS.
There are two kinds of a fool:
One says +This is old and therefor good+
The other says +This is new and therefor better+
Paul C. George
Date: Thu Jun 30, 1994 11:43 am PST
Subject: Re: Albus and Powers
[Paul George 940630 10:30]
>From Tom Bourbon [940629.1731]
> In contrast, Albus began the fourth article in his series (which wastitled, "A model of the brain for robot control") by saying, "The essence of ahierarchy is that control is top-down. The ultimate choices are made at thetop, and the goals selected at this level are decomposed into action as theyfilter down through the various levels of the hierarchy." That's about as faras you can get from the idea about a control hierarchy in PCT; top-downhierarchies of that kind don't work as models for living systems.
Note that the article I cited is the only Albus work I have read. It wascirculated as a theoretical basis for a hierarchical distributed object controlsystem architecture the company is working on.
Aldus's thinking appears to have evolved since 1979 (one would hope). Thearticle is not oriented around top down control as you appear to think of it.Indeed, it is far more bottom up in nature, and focused on coordination.Control policies or goals _may_ flow down the hierarchy, but summarized dataand model results (error signals) pass up. Both also move laterally in thenetwork. It is not gospel, just a nice description of the workings of controlsystem networks (once you get past the definitions and philosophy). I will notdefend it as TRUTH. As Martin commented, his definition of language seems wayoff base to me, and there were a number of other things that caused raisedeyebrows. I would encourage you to read the article and comment on the wheatand chaff.
BTW, I do not see how PCT could refer to a control system _Hierarchy_ isthere is no kind of command relationship, or hierarchical filtering andsummarization of data. That is part of the definition of 'hierarchy'. The quoteyou cite is fundamentally accurate. However, control systems need not be_strictly_ hierarchical. And equally clearly living systems do involvehierarchical structures.
If you deny the existence of hierarchical control, perhaps you should use'tree network' instead. I think it is more likely that you are having a NIHreaction, focusing on minor distinctions as the most important, for the purposeof differentiation. PCT aficionados do not have a monopoly on understanding,even if they may have produced working models.
Paul C. George
Date: Thu Jun 30, 1994 11:48 am PST
Subject: Re: Albus and Powers
Tom Bourbon [940630.1201]
>[Paul George 940630 10:30]
>>From Tom Bourbon [940629.1731]
>> In contrast, Albus began the fourth article in his series (whichwas titled, "A model of the brain for robot control") by saying, "The essenceof a hierarchy is that control is top-down. The ultimate choices are made atthe top, and the goals selected at this level are decomposed into action asthey filter down through the various levels of the hierarchy." That's about asfar as you can get from the idea about a control hierarchy in PCT; top-downhierarchies of that kind don't work as models for living systems.
> Note that the article I cited is the only Albus work I have read. Itwas circulated as a theoretical basis for a hierarchical distributed objectcontrol system architecture the company is working on.
Understood. But recall that you appeared on this net with an announcementthat, ". . . you may be using too simplistic a view of a control system, whichis biasing your discussion (in fairness I haven't had access to thebooks/papers on the subject . . .", followed by, "Sophisticated control systems[TB: by implication, PCT is unsophisticated] don't use reference variables(e.g. setpoints or alarm thresholds) they use reference _models_ (reflected incontrol logic). [TB: PCT uses all of these things you say are not used bysophisticated models.] You concluded with:
=================
For a good exposition of this model I suggest:
"Outline for a Theory of Intelligence" James S. Albus IEEE Transactions onSystems, Man, and Cybernetics, vol 21 #3, May/June 1991, p 473-509. IEEE log #9042583
==============
which looked like a pretty strong endorsement of Albus as an example of asophisticated modern model of control.
I hope you will look back over your initial posts and see if theirappearances might have something to do with what you see as defensiveness onour part. We _do_ have a history of coming under direct and heavy "attack" bypeople taking positions close to those you stated in your initial posts. But asyou can see, we have been eager to ask questions and learn more about yourideas, while at the same time admittedly putting up a semblance of a defenseagainst what looked at least a little bit like another attack.
I hope you will find the interest and the time to reply to some of thespecific questions I addressed to you.
> Aldus's thinking appears to have evolved since 1979 (one would hope).The article is not oriented around top down control as you appear to think ofit.
In fact, in his series in 1979, Albus included some very vague and generalideas about other kinds of connections in his "hierarchy," but he neverdeveloped a model -- just a set of conjectures on how organisms (brains)produce specified outputs. His model was, and from what you say of the newerarticle it remains, a model for the production of predetermined outputs. I'llneed to look at the 1991 article before I say much more about that,though.
> Indeed, it is far more bottom up in nature, and focused oncoordination. Control policies or goals _may_ flow down the hierarchy, butsummarized data and model results (error signals) pass up.
The error signals go up? From what? To what? What role(s) do they play?That's the opposite direction from the flow of error signals in the elementalPCT model. (A question and an observation; nothing defensive intended.)
> Both also move laterally in the network.
From where to where, with what effects? Are there quantitativeimplementations of this model? Does it actually behave, or is it only requiredto sound plausible? (I hope you don't find my questions "defensive;" I reallydo want to know.)
> It is not gospel, just a nice description of the workings of controlsystem networks (once you get past the definitions and philosophy).
This comment seems to suggest that the Albus model really _hasn't_ changedall that much since 1979, at least not with regard to its nature (a descriptive"model") and its role (sounding plausible -- nice). Again, I hope you don'tfind my comments defensive; I'm merely stating my impressions drawn from the1979 articles and your description of how you use the ideas from the 1991article.
> BTW, I do not see how PCT could refer to a control system _Hierarchy_is there is no kind of command relationship, or hierarchical filtering andsummarization of data. That is part of the definition of 'hierarchy'.
Ah, that's why you might want to read some of the PCT literature before yousay much more about the PCT model. You see, you've identified one of theamazing things about a hierarchy of simple control loops, each of whichcontrols only its own perceptual signals -- such a hierarchy accomplishes featsthought impossible by those who are familiar only with traditional notionsabout hierarchies. Look at some of the literature, then let us know what youthink.
By the way, if you do look at the PCT model, you'll find there are _no_"commands," but that there is indeed a hierarchical "summarization" ofperceptual signals. Living systems have their own "definition" of hierarchyand it doesn't much resemble the more common definitions in the behavioral andlife sciences.
> If you deny the existence of hierarchical control, perhaps you shoulduse 'tree network' instead.
But we don't deny the existence of hierarchical control, only of "command"hierarchies that control outputs. Living systems hierarchically control theirown _inputs._ "Tree network" doesn't do, in that case; it's a full blownhierarchy.
> I think it is more likely that you are having a NIH reaction, focusingon minor distinctions as the most important, for the purpose ofdifferentiation.
Not really. Read the literature; test the models; then let us know whatyou think about this idea.
> PCT aficionados do not have a monopoly on understanding, even if theymay have produced working models.
That's for sure! But we are certainly in a position to reply to anyone whotells us our model is too simplistic and unsophisticated to work, or that itcannot really be of the kind we say, or that it is equivalent to word-modelsthat have never been required to prove their worth by actually behaving inhowever simple a setting. And we will certainly reply if the someone also tellsus he or she hasn't read any of our work or tested any of our models. That'shardly the NIH reaction.
Happy reading. ;-))
I'm serious. You have gone to the trouble to emerge out of the pool oflurkers and engage in some discussion. I'm sure all of us who have been on thenet for a while would enjoy seeing your thoughts about the PCT model after youhave learned a bit more about it.
Later, Tom
Date: Thu Jun 30, 1994 3:56 pm PST
Subject: Back in service; a few comments
[From Bill Powers (940630.1330 MDT)]
Paul George:
Welcome to CSG-L! I echo Tom Bourbon's comments with, perhaps, a littlemore diplomacy: I, too, think it would be a good idea if you were to becomefamiliar with the PCT model before drawing any important conclusions aboutit.
It's important to understand the control-of-input idea, and why control ofoutput doesn't work for real control systems in real (i.e., variable)environments. Human beings NEVER compute outputs: doing so would be futile. Tosee why, just imagine driving your car by planning the motor forces you wouldapply to the steering wheel while driving to work. When we speak of planning,what we're always speaking of are planned perceptual consequences, not thedetailed actions that will bring them about. When you plan a route for drivingto the store, the route isn't an output, but a series of perceptions you willachieve by whatever means is required when you get to the appropriate part ofthe plan. There's no way to predict exactly what output will be required toachieve a given perceptual result. Fortunately, control systems base theiractions on error, not planned outputs, and so can produce the right action overa wide range of external conditions.
> BTW, I do not see how PCT could refer to a control system _Hierarchy_if there is no kind of command relationship, or hierarchical filtering andsummarization of data. That is part of the definition of 'hierarchy'.
The "command" relationship in PCT is not one of commanding actions, but ofspecifying the states of perceptual signals that are under control by lowersystems. This specification is done by sending a reference signal to a lowersystem which is like a sample of the lower perceptual signal in the desiredstate: make it look (feel, sound, taste) like _this_. This is quite differentfrom the way other theorists have treated the relationship between levels. Youreally need to read BCP to see how this works, and particularly how it can workat many levels simultaneously.
Best to all, Bill P.
Date: Fri Jul 01, 1994 8:24 am PST
Subject: Re: PCT models
[Paul George 940701 10:00}
A last comment on this thread. Thanks to all for the welcomes. From mypoint of view, what you are discussing in PCT is not particularly strange. Thepatterns are familiar, if uncommon in the psychological and medical arenas.There seems to be mostly a communication and terminology problem, at least withTom. I am mildly amused when I am told I don't understand PCT, immediatelyfollowed in another post by an example of what I mentioned.
>[Bill Powers (940630.1330 MDT)]
> The "command" relationship in PCT is not one of commanding actions,but of specifying the states of perceptual signals that are under control bylower systems. This specification is done by sending a reference signal to alower system which is like a sample of the lower perceptual signal in thedesired state: make it look (feel, sound, taste) like _this_. This is quitedifferent from the way other theorists have treated the relationship betweenlevels. You really need to read BCP to see how this works, and particularly howit can work at many levels simultaneously.
<Bob Clark (940630.1450 EDT)>
> [For convenience, here is a quick summary of HPCT concepts: 1)"control of perception" and 2) a "hierarchy of levels of control" 2a)"hierarchy" is defined as a relation between "levels" in which "higher levels"achieve their objectives by selecting and activating goals, "reference levels,"for lower level systems.]
Hey guys, this is exactly how distributed control systems and softwaresystems are architected, _and_ is the architecture described by Albus in thearticle I cited. It's not particularly revolutionary in the engineering field.In a sense the model has been validated not by simulation, but rather byconstruction. Of course that just means the architecture works, not that is ofnecessity the mechanism used by nature. However, your low level simulations doappear (from discussion and description) to be well on the way to 'proving' thePCT approach at the neuro/muscular and sensory perception level.
>[From Bill Powers (940630.1330 MDT)]
> It's important to understand the control-of-input idea, and whycontrol of output doesn't work for real control systems in real (i.e.,variable) environments. Human beings NEVER compute outputs: doing so would befutile.
> From Tom Bourbon [940630.1645] {offline communication} The PCT modeldoes not control its outputs. Neither do living things. Again, your commentsare uninformed, as they must be, given that you have read none of ourwork.
Perhaps there is confusion about the term 'output'. Control software (orfirmware) produce outputs (signals) to actuator mechanisms which produce the'environmental' output. I know of no one in engineering who suggests that thecontrol system directly influences the external environment. If I try to catcha ball, I do not control eye tracking or arm/finger muscles. When I attempt to,I miss.
If I am constructing control system software, I have sensors and datacollectors that sample and translate (A to D) the sensor output. Even thesensor does not directly provide the environment's condition, just some kind oftransducer reading. Similarly if I wish to move a robot arm, the control systemmight output a command to a particular joint to move to a particularcoordinate. Some mechanism translates that to providing a current or fluid flowfor a period of time at certain points (this may be deemed another control nodeif you wish). This translation is based upon some model of what _should_ happento the arm when this occurs, in terms of position or velocity. However, ifthere is no power or hydraulic fluid, or if the arm is blocked, nothingactually happens, while the control system 'thinks' it has. This is why we haveother sensors to detect actual motion, though they too can fail or bespoofed.
Further, the control software actually has objects that represent thesensors and arm components being controlled. It is these 'simulations' or'models' of the real world entities with which the control logic interacts.This, it appears, is PCT's controlled perceptions. For process control systemsthere are models of the physical process being controlled, and of the controlsystem itself. One of the hard engineering problems is making sure this modelstays synched with reality. As the old saw goes "the map is not theterritory".
Enough for now. You guys aren't as alone as you might think, thoughperhaps a bit insular. Other domains deal with the same kind of issues, andhave come up with solutions and approaches which may apply.
Paul C. George
Date: Fri Jul 01, 1994 12:08 pm PST
Subject: Re: PCT models
[From Bill Powers (940701.1100 MDT)] Paul George (940701.1000)
It's fascinating how one can read into expositions what one expects to findthere. You seem convinced that PCT is just the same thing that controlengineers are already doing, yet in explaining back to us what they do you aredescribing something very different. Not your fault; this is a common propertyof human beings. In fact, a few control engineers have told us that the PCTapproach in control engineering is new to them and suggests a very differentway to design control systems -- even though the equations might be exactly thesame, as they sometimes are!
Maybe I'm misunderstanding you, but here is what I based my commentson:
> Perhaps there is confusion about the term 'output'. Control software(or firmware) produce outputs (signals) to actuator mechanisms which producethe 'environmental' output. I know of no one in engineering who suggests thatthe control system directly influences the external environment. If I try tocatch a ball, I do not control eye tracking or arm/finger muscles. When Iattempt to, I miss.
Good guess, but not the point we are making. The actual point comes up inthe following:
> If I am constructing control system software, I have sensors and datacollectors that sample and translate (A to D) the sensor output. Even thesensor does not directly provide the environment's condition, just some kind oftransducer reading. Similarly if I wish to move a robot arm, the control systemmight output a command to a particular joint to move to a particularcoordinate. Some mechanism translates that to providing a current or fluid flowfor a period of time at certain points (this may be deemed another control nodeif you wish). This translation is based upon some model of what _should_ happento the arm when this occurs, in terms of position or velocity.
This is exactly the conventional control-of-output conception of controlthat we are deviating from. In the conventional engineering view, the problemis to cause some objective effect to appear in the environment (like moving agripper to a preselected objective position). One way to do that is to computethe outputs that must be generated in order to have that effect, includinginverse kinematics and dynamics and any effects from disturbances that might beanticipated, and then to send the output signals to the transducers to producethe effect. Feedback then is used largely to trim the dynamic characteristicsof the system for stability and to provide resistance to disturbances.
The PCT approach starts at a different point in the loop: not the objectiveeffect, but the perceptual representation of it. This means that feedback ismore than just a way to trim the system response and take care of disturbances:it is the only way a living organism can know ANYTHING about the effects itsoutputs are having in the external world. When a living control systemcontrols, all it can control is the perceptual representation; it has noauxiliary channels through which it can know the objective effects of what itsoutputs are doing to the world.
This means that when, as engineers, we design a system from the PCTstandpoint, the first step is to create a perceptual signal that reflects thestate of the external variable to be controlled. As designers we have theadvantage over the organism in that we can preselect what the external variableis, but in the spirit of PCT we follow the same strategy to which the organismis limited: we set up a perception that is the real controlled variable. Infact, what is perceived is what is controlled; if we want to make sure thatsome objective variable is controlled, we have to make sure that the perceptionaccurately reflects its state. This is not quite the same problem that theorganism has, for from the organism's point of view, what is perceived isalways what is controlled, and the organism usually doesn't know what externalsituation corresponds to control of that perception.
Given that the perceptual signal is what we want to control, we now have asimple problem on the input side. The reference signal specifies the desiredstate of the perceptual signal, and the comparator detects the error. If theperceptual signal exactly matches the reference signal, the externalcounterpart of the perceptual signal -- which the engineer probably calls thecontrolled variable -- will be in a particular state. All of the design problemthat remains is then that of devising an output function that will (a) createthe necessary environmental effects to complete the control loop, and (b)introduce the dynamic functions needed to achieve stability in the presence ofdisturbances and changes in the reference signal.
In a PCT control model, two features of the standard engineering designsare missing. First, there is no model of the environment contained in thecontrol system (although of course there must be one in the simulation as awhole). Second, there are no inverse calculations, either of kinematics or ofdynamics. The system is set up so that the error signal is translated intooutput effects that make the error smaller; by suitable choices of controlledvariables (including hierarchical relationships), this is sufficient toguarantee tight control and stability.
As an example, consider "Little Man Version 2," a model of limb control Ihave been working on for a couple of years. I took my design straight fromneurophysiology, using a model of the muscle and of the tendon and stretchreflexes that is a simple and literal rendition of how the neuromotor system isactually wired. If there's anything clever in this design, give the credit tonature.
The lowest-level control loop is the tendon reflex, which senses the forceapplied by a muscle through a tendon to an attachment to a bone. The force issensed directly as a neural signal which is compared with a reference signal,and the error is amplified to shorten or relax muscle fibers. Shorteningstretches a series elastic component to produce the force that is sensed,closing that loop. This system causes the applied force to track the referencesignal.
The second loop is a rate-control loop, and the third is a position-control loop. The position error determines the rate reference signal, and therate error determines the force reference signal -- a three-level hierarchy,when decomposed this way. In the real system, the two higher loops are combinedinto one, with the rate control being combined with position control anddetermining the damping of the system.
What's interesting about this model is that stability is achieved withoutany special attention. With acceleration, velocity, and position all underspecific and tight hierarchical control, the arm dynamics (which are includedin the model of the system's environment) are almost perfectly compensated,without any need for inverse dynamic calculations, Jacobians, or all thatstuff. Of course a real control engineer could probably work out how thisdesign actually takes the required inverse calculations into account, but infact the control system model itself does no such calculations. The dynamicalequations appear only in computing the response of the arm to torques appliedat its joints, which is really a model of the environment, not of the controlsystem.
There is nothing in this model that could be construed as calculating thedesired effect of an output and then producing the signals that will createthat desired effect. The output calculations, in their entirety, consist ofleaky integrations. What has struck some control engineers as interesting aboutthis model is its extreme simplicity in comparison with other approaches. Allsix control systems involved in controlling three degrees of freedom in the armare expressed as no more than 75 lines of rather loose C code, with no matrixoperations or other shortcuts. The other 5000 lines of code are concerned withgraphic presentation, facilities for parameter adjustments and test modes, andmodeling the physics of the arm.
There is one sense in which it can be said that the Little Man modelincorporates a full-time running model of the environment. As far as thecontrol systems are concerned, the perceptual signals ARE the environment. Eachperceptual signal is a different function of environmental variables, and sorepresents the state of the variables as the variables in a running modelrepresent the state of something else. We could say that the model is there --but from the standpoint of the control system, it is simply "reality."
In the final analysis, PCTers and control engineers are really talkingabout the same thing. The main difference is in how one parses the systembefore modeling it. There's nothing wrong with the way control engineers goabout their business, although I sometimes think that they have overcomplicatedsome problems, but the way they go about their business isn't the rightapproach for modeling organisms, which is what PCT is about. There is nofriendly engineer standing by to tell an organism what it is really doing toits environment when it controls its own perceptions. Whatever an organism canfigure out about the external world has to be done entirely through informationcarried in its perceptual signals, which are its feedback signals, andfunctions of those perceptions, and functions of those functions. It's asthough the organism's brain lived in a control room lined with meters andlevers, but the meters carry no labels showing what their readings representand the levers have no labels describing what they do to the world outside thecontrol room.
I have always hoped that real control engineers would lend their expertiseto expanding the PCT approach. Now some of them are actually trying it out, andI hope to see more sophisticated analyses than I can produce. The first hurdlehas always been to convince control engineers that there IS another approach,and even though it is mathematically equivalent to standard ones, it is alsovery different -- and simpler to implement -- in many regards. It's moreimportant to understand the differences than to see the similarities.
Best, Bill P.
Date: Fri Jul 01, 1994 2:18 pm PST
Subject: Re: PCT models
From Tom Bourbon [940701.1601] >[Paul George 940701 10:00}
> A last comment on this thread.
I hope it isn't concluded. Much remains to be discussed andclarified.
For example, in my first post or two, I asked you how you go aboutassessing models and theories -- which criteria and procedures do you usebefore you decide a theory or model is "good" or that it "works?" For onething, your answers to those questions would help me, and I'm sure many otherson this net, to prepare appropriate replies to your messages. A general answerwould help, but a specific example of how you used those criteria andprocedures to assess Albus's model would be even better. I hope yourannouncement of the end of this thread does not preclude an answer to thosequestions.
> Thanks to all for the welcomes.
Thanks to you, for breaking your silence.
> From my point of view, what you are discussing in PCT is notparticularly strange. The patterns are familiar, if uncommon in thepsychological and medical arenas. There seems to be mostly a communication andterminology problem, at least with Tom. I am mildly amused when I am told Idon't understand PCT, immediately followed in another post by an example ofwhat I mentioned.
I hope there isn't a serious communication problem between us. My initialresponses to you might have seemed abrupt -- perhaps that is my style when I amtrying to catch up on a stack of accumulated mail, answering piece by piece,rather than waiting until I've read the entire pile. My intent was to suggestthat your impressions about PCT, as it is named and discussed on this net,might not be accurate. I was not trying to imply either malice or stupidity, oranything akin to them, on your part. I _was_ reminding you of your ownadmission (which I respect) that you had not read material on PCT, hence, yourideas about it, and your attempts to compare it to other ideas (Albus's theory,for example) might suffer by virtue of your lack of exposure to PCT. I believethe replies you received from participants on this net _all_ have been civiland welcoming, and that all have emphasized the idea that an examination of thematerial on PCT might be important for you, if you wish to enter the discussionof the phenomenon of control and of PCT.
I'm puzzled by your statement that you are "mildly amused" to read that youare uninformed about PCT, as it is discussed on this net. You did say you hadnever read anything about it. In my case, I'll freely admit that I'm ignorantof (uninformed about) the actual contents of the Albus paper in 1991. I believemy ignorance and state of uninformedness on the paper disqualify me from making_any_ pronouncements about his theory or his model, circa 1991, although I doknow his ideas circa 1979 and they do not seem to differ dramatically from whatyou have said about the more recent article. But I will make no furthercomments about his "model" or about your interpretations of it until after Ihave read it this weekend.
Uninformedness and ignorance are merely descriptors of a state of absenceof knowledge about a particular topic. The terms do not imply a value judgmentabout the person who is ignorant or uninformed on that topic.
Later, Tom
Date: Fri Jul 01, 1994 4:03 pm PST
Subject: Re: PCT models
[Paul George 940701 15:00 EDT]
OK I'll bite. Principae Mechanica is not a prerequisite for catching aball.
[From Bill Powers (940701.1100 MDT)]
> It's fascinating how one can read into expositions what one expects tofind there.
> This is exactly the conventional control-of-output conception ofcontrol that we are deviating from. In the conventional engineering view, theproblem is to cause some objective effect to appear in the environment
It seems to work on both sides. You conveniently ignore the laterdescription of process control software and intermediary steps which matchexactly your discussion.
> This means that when, as engineers, we design a system from the PCTstandpoint, the first step is to create a perceptual signal that reflects thestate of the external variable to be controlled. {snip}
> What's interesting about this model is that stability is achievedwithout any special attention. With acceleration, velocity, and position allunder specific and tight hierarchical control, the arm dynamics (which areincluded in the model of the system's environment) are almost perfectlycompensated, without any need for inverse dynamic calculations, Jacobians, orall that stuff. ... The dynamical equations appear only in computing theresponse of the arm to torques applied at its joints, which is really a modelof the environment, not of the control system. {snip} There is nothing in thismodel that could be construed as calculating the desired effect of an outputand then producing the signals that will create that desired effect. The outputcalculations, in their entirety, consist of leaky integrations. What has strucksome control engineers as interesting about this model is its extremesimplicity in comparison with other approaches.
I said the control system defined a desired result - "move to particularcoordinates". Such result is determined from the inputs. The control node hasaccess to a 'perception' of the external world entities which are the mirroredobjects (we call this a 'process centric view'. Process means something like'make chemical with a certain recipe'. 'Self' is one of those entities). It isa set of instance variables. There also exist reference variables to which thisobject 'state' is compared (albeit by changing formulae and sets of 'controlledvariables').
I said a 'mechanism' exists to generate outputs to bring the 'perceived'state in line with the 'reference' state. I said nothing about reversekinematics, nor of feedback, nor of computing outputs _within the controlnode_, though this is a conventional mechanism for robot control. Your errorcorrection method in the 'Little Man' simulation is just another (though moreelegant) mechanism. I would like however to see it directly applied to a robotarm.
> In fact, what is perceived is what is controlled; if we want to makesure that some objective variable is controlled, we have to make sure that theperception accurately reflects its state.
Yes, as I said, the control signals (error signal) are sent to the mirroredobject which reflects the real world entity. That object then has methods toaffect the outside world through some chain of transducers and actuators.Measurement theory states that a measured value need to have a knownrelationship (not necessarily liner) to the thing being monitored, within aparticular range, such that it is useful for prediction or control. A 'model'of sorts is needed to determine what output will cause a change of thecontrolled variable in the desired direction to diminish the error signal. Themodel _can_ of course be a simple mathematical function. It could however be acomplex hierarchical finite state machine (which I think _could_ be implementedas a network of simple control loops).
> The other 5000 lines of code are concerned with graphic presentation,facilities for parameter adjustments and test modes, and modeling the physicsof the arm.
These also must exist in a control system at above level 2 (0=AtoDconversion, 1 = PLC, 2=local net or process, 3=plant, 4=enterprise). At thelevel of nerve and muscle your mechanism works well, and probably does map toreality. However at higher levels _everything_ is virtual. You are usually'controlling' lower level's reference variables and comparison algorithms(though you may have higher level controlled variables). They are also morefocused on planning or predicting _future_ actions (or results) based uponextrapolation of current trends. The key idea is that granularity of data(input & output signals, reference & control variables) and timehorizon increase as you go up the hierarchy. In my terms, your position, rate,and neuro/muscular loops are on the same level, though with a directionalcontrol flow. Note that in the (3D) Albus(perhaps giving him too much credit)architecture, nodes on the same level interact directly.
BTW - have you tried to cause the object being tracked in your simulationto change velocity in 3 space randomly? I suspect the little man would havetrouble pointing at the target. You might find that an environmental modelincorporating the physics of motion was needed. To make the problem harder, weobserve that trying to 'catch' objects under water, or in lower gravity, ordifferent coriolis force, is difficult and requires re-learning because thingsdon't move as we expect. What happens if we make the little man observe in amirror? How do you correct your output signals when the error signal doesn'tdiminish as expected? {My apologies if these are already dealt with, I haven'tyet received Dag's package}
How about at the level of verbal communication? I construct some set ofstatements to influence your behavior. This is based both upon my perception ofyour response, my understanding of language, and my predictions of yourreaction. This seems a bit beyond a simple error signal as there isconsiderable lag and indirection. It does seem based upon some set ofpredictive world models, or if you prefer, simulations. I compare the resultsof my model to my perceptions, and modify the model accordingly. (here controlsystem = conscious & unconscious mind). However, I can view these models ascontrol networks themselves encapsulated within a control system.
Hope there is a spark of something new here, If nothing else that differentviews or terminology can apply to the same thing. See any responsesTuesday
Paul George
Date: Fri Jul 01, 1994 7:39 pm PST
Subject: Re: PCT models
<[Bill Leach 940701.21:58 EST] >[Paul George 940701 15:00EDT]
Paul, with Bill in the discussion I am reluctant to comment since hispostings are so concise and usually outstanding... however, I will anyway iffor no other reason than that typically I learn something.
Control system engineers should understand fundamental PCT quite rapidly.Understanding the deep implications of PCT might take quite a bit more time andeffort.
For control systems engineers, it is important to remember that:
1. PCT is closed loop negative feedback control only. No open loopstuff.
2. The control system IS everything. There is no possible exact independentcheck of performance (at least for higher levels).
3. The control system IS adaptive.
4. The control system ALWAYS sets it's own references.
> BTW - have you tried to cause the object being tracked in yoursimulation to change velocity in 3 space randomly? I suspect the little manwould have trouble pointing at the target.
No doubt however, you provide the answer to your own question:
> ..., we observe that trying to 'catch' objects under water, or inlower gravity, or different coriolis force, is difficult and requiresre-learning because things don't move as we expect.
It is quite likely that humans don't do any "calculations" for issues ofmotion. That many have the ability to recognize certain types of motion isitself probably a learned ability that is independent of other control systemfunctions.
-bill
Date: Sat Jul 02, 1994 4:44 am PST
Subject: Re: PCT models
[From Bill Powers (940702.0400 MDT)]
[Returned from conference in Wales] The internal clock is slowly creepingback into synch; good thing I don't have to go to work.
Paul George (940701.1500 EDT)--
We may slowly be converging to a common description of the controlphenomenon in behavior. You should realize that when I began this work some 40years ago, my target audience was in the life sciences, and I simply hoped tocommunicate what control engineers knew to people working (unknowingly) withliving control systems. So I've always hoped to see real control-systemengineers getting into this act. I would be quite alarmed to find that PCT andcontrol engineering had nothing in common!
I have not really ignored your comments on process control. My problem issimply that in the language you use to describe control, you use terminologythat is like what others have used in expressing a theory of control that isvery different from PCT. It isn't easy to see through language to theunderlying concepts behind it; the language may suggest divergences where thereare none, but it can also suggest agreement where there is actually adivergence. So we just have to slog through this and make sure we are talkingabout the same things. Words don't always convey to a listener the meaningsthat the speaker had in mind.
Here is an example:
> I said the control system defined a desired result - "move toparticular coordinates".
As I read this sentence, it says that the control system moves something toparticular external spatial coordinates, in an objective coordinate system.This is not how we would describe the situation for an organism in PCT. What wewould say is that the control system acts on the world in such a way thatperceptions of position -- i.e., signals representing position in approximatelyorthogonal ways -- come to particular values. The measure of position is not ameter-stick reading obtained by an external observer in the environment of thecontrol system, but a set of perceptual signals existing inside the controlsystem. So even at this level (in fact, at every level), the world beingcontrolled is, as you say, "virtual."
You then add
> Such a result is determined from the inputs.
Here's the seeming problem. As you describe the situation, there is firstan objective position, produced by actions of the effectors. Then this positionis represented in the system by signals arising from its input sensors. Fromthis point of view, it is possible for the perceptual representation ofposition to be incorrect or distorted. But from the organism's point of view,the perceptions ARE the position; what is controlled is not the objectiveposition, but the perception. Given that the perceptions in each coordinate aremaintained at particular values, the objective position that corresponds tothose values is simply the inverse input function of the perceptual signals. Ingeneral, when the reference signals for position are varied, the resultingposition in external coordinates will change along curved lines, the curvaturebeing determined by the nonlinearity of the sensing apparatus (for example, the"horopter" of visual three-dimensional space). So the space in which theorganism controls position does not coincide with the Cartesian or polarcoordinate systems that an external observer might use to characterizebehavior.
You say
> The control node has access to a 'perception' of the external worldentities which are the mirrored objects (we call this a 'process centric view'.Process means something like 'make chemical with a certain recipe'.
Good: the process-centric view is then what we refer to as control ofperception. However, the strictly correct way to describe the process wouldthen be 'make the perception of a chemical appear by manipulating perceptionsof a certain recipe.' In the purely process-centric view, the laws of chemistryand the objective consequences of carrying out a given recipe are known to thecontrolling system only as perceptual signals; what is actually happening maybe different by a small or large amount. From the standpoint of a controllingprocess, the objective world is known ONLY through representations in signalsgenerated by its input transducers, and computations performed with thosesignals as arguments. Furthermore, what is controlled can be only therepresentations; the external correlates of those representations may not bewhat an external observer assumes.
I don't follow this: did you say what you mean?
> There also exist reference variables to which this object 'state' iscompared (albeit by changing formulae and sets of 'controlledvariables').
I understand comparing one signal against another, but how is such acomparison done by "changing formulas and sets of controlled variables?" Didyou really mean that the _comparison_ is done by these means?
> I said a 'mechanism' exists to generate outputs to bring the'perceived' state in line with the 'reference' state. I said nothing aboutreverse kinematics, nor of feedback, nor of computing outputs _within thecontrol node_, though this is a conventional mechanism for robotcontrol.
Sorry if I misread you. What alerted me to a possible disagreement was thefact that others have proposed "mechanisms" designed to bring the objectivevariable, and thus the perception of it, TO a reference state rather thansimply TOWARD it. This interpretation has turned up in many places, among themcognitive psychology and the motor control literature. The underlying modelrelies on assessing the situation, then computing what action is required tobring an external variable to a goal-state, and finally executing the programthat will accomplish the pre-planned result. That is how the need for computinginverse kinematics and dynamics arises. Where feedback is involved, it occursafter the pre-planned movement, and serves to inform the system about theresults, so that the next movement can be planned. This type of control loopmight appear in some circumstances, but is far from adequate as a general modeland is not a particularly effective mode of control. It is necessarily veryslow because of all the calculations that need to be done to make the resultscome anywhere near the desired goal, and it relies on the absence ofdisturbances occurring between the initial assessment and the final executionof the plan.
I now take it that this is NOT the model you are speaking of.
> Your error correction method in the 'Little Man' simulation is justanother (though more elegant) mechanism. I would like however to see itdirectly applied to a robot arm.
Thanks. I would like to see it applied that way, too. In fact, thatapplication might impress grant-givers enough for them to supply the fundsnecessary to develop that application. Catch-22. In the meantime, the best Ican do is a careful simulation.
> ... as I said, the control signals (error signal) are sent to themirrored object which reflects the real world entity. That object then hasmethods to affect the outside world through some chain of transducers andactuators.
If I understand this rather terse statement as you intend it, this seems tobe somewhat like the concept of hierarchical control in PCT. The error signalfrom a higher-level control system, after passing through an output function,fans out to enter a set of lower-level control systems each of which isconcerned with controlling one dimension of the perceived world. However, it'snot clear in what sense your "mirrored object" reflects the real world entity.In hierarchical PCT (HPCT), all that is represented explicitly, in the form ofsignals, is the value of a variable: the properties of the external world arenot represented as signals. We have toyed with the idea of actual analog modelsof the external world being used by the control system, largely to fill inmissing information, but so far have not come across a case that requires this(not to say that such cases will not be found).
The lower-level "object" in HPCT consists of a lower level of controlsystems, plus the environment with which they interact. In the absence of aliteral model, that lower-level environment serves as its own model; that is,it responds to output signals from the control system by altering input signalsin the control system. Any mirroring of the properties of the external worldwould be contained in the organization of the perceptual (input) and outputfunctions of the control system -- for example, the appearance in the outputfunction of the complex conjugate of external dynamics (to use a term from theold frequency-domain days).
You seem to agree with this when you say
> A 'model' of sorts is needed to determine what output will cause achange of the controlled variable in the desired direction to diminish theerror signal.
I think that a "model of sorts" is an improvement over the more literalconcept of a complete analog model of the environment, which some have proposedbut doesn't seem practical to me.
Do I detect a little influence here from OOPS? One has to beware of lettingprogramming languages dictate theories of behavior!
> These also must exist in a control system at above level 2 (0=AtoDconversion, 1 = PLC, 2=local net or process, 3=plant, 4=enterprise).
This sounds reminiscent of Albus. We use an entirely different set ofconjectures about the nature of higher-level control processes. Albus'definitions of levels are not tied very well to real behavior and experience --they're more like an engineer's assessment of an objective problem in plantcontrol.
I've been concerned with trying to identify broad types of variables thatpeople seem actually to control, relating them as nearly as possible toneurophysiology and direct experience. What is referred to above as "A-Dconversion" occurs in the HPCT model at the 7th proposed level of control. I donot consider that the basic sensing process is an A-D conversion: it is aconversion from stimulus amplitude to impulse frequency, which is still ananalog measure. The conversion from analog processes to digital processes, Ihave guessed, occurs where we create categories of perception out of thecontinuum of analog perceptions at lower levels. This is the level wheresymbols come into existence, representing either-or classifications rather thancontinuous processes. Above that level we begin to have what is usuallydescribed as a finite-state machine -- a machine that can carry out discretesequences of acts and that reasons in terms of arithmetic, logic, grammar, andother such program-like rule-driven methods.
I'm sure that within the levels currently defined under HPCT, we couldaccount for the things referred to in your Albus-like list above, but obviouslywe are slicing the pie along different planes.
> At the level of nerve and muscle your mechanism works well, andprobably does map to reality. However at higher levels _everything_ isvirtual.
Actually, in the HPCT model, everything is virtual at all levels, as I saidearlier in this post. It isn't that the perceptual hierarchy "maps to reality"-- it _creates_ reality as we know it, out of the raw input of stimulusintensities at the lowest level. As a friend of mine said, the organism isn'tthe black box; the environment is. All that an organism can know of realitycomes from seeing what happens when it emits output signals into the worldbeyond the senses, and observes the resulting changes in its own perceptualsignals. In our modeling, we use physical models of the external world derivedby the same means, and look for correspondences between that physical model andthe world we experience as reality. These worlds are not very similar, althoughthere are many points of contact. The ultimate criterion is not what thephysical models say, but what experience tells us directly. Physics is validonly to the extent that it correctly predicts experience (as, I should add, itdoes exceedingly well).
> You are usually 'controlling' lower level's reference variables andcomparison algorithms (though you may have higher level controlledvariables).
I would say "varying" rather than "controlling" lower-level referencevariables. What we control are input signals; we do this by _varying_ outputs(reference signals for lower level systems) as required by variations inenvironmental conditions (disturbances).
I should mention that the PCT model works strictly with scalar perceptualvariables; vector inputs are reduced by higher-level input functions to thevalues of one-dimensional variables such as distance and angle. The result isactually consistent, pretty much, with models expressed in matrix notation, themain difference being that in the HPCT model all the matrices are presented ascompletely expanded. When you think about it, that has to happen in any realsystem anyway: the matrix notation is just a way of handling a lot of moredetailed, and simpler, processes in a convenient way. The real system still hasto do every individual operation implied by the matrix operations (just as acomputer must).
The reason I mention this is that when control systems are seen as workingwith scalar perceptions only, there are no complex "comparison algorithms."Comparison is simply subtraction. All the complexity that would appear in othermodeling approaches happens in the perceptual input functions at each level,which apply algorithms that extract, in parallel, many different scalarmeasures from the common set of lower-level perceptual signals. The generalrule in HPCT is "one perception, one control system." Complex control isachieved by many such systems operating in parallel at the same level. I don'tknow how long this concept will survive; right now it seems to work pretty wellin terms of experimental tests.
> They are also more focused on planning or predicting _future_ actions(or results) based upon extrapolation of current trends.
I agree that there are cases in which planning and extrapolating do occur,but generally the kind of control you get from this approach is poor. Realenvironments are too variable to permit much more than statistical success fromthis method. Where it's possible, and I think it is possible much more oftenthan some people seem to think, a better approach to controlling is to get thecomponent variables under individual control at a lower level, then simplycontrol higher-level derived variables relative to specific reference levels:don't try to _predict_ the future, _control_ it. Where you can do that, itworks a hell of a lot better than making predictions and plans. A compromisemethod is contingency planning, which is much more like real control. Insteadof deciding what you are going to do in complete detail, you set up controlsystems to handle a range of possible occurrences, without trying to predictwhich of them will occur.
> The key idea is that granularity of data (input & output signals,reference & control variables) and time horizon increase as you go up thehierarchy.
Well, that may be true, but it's not a very useful concept for makingmodels. It's sort of like Ashby's "Law of Requisite Variety," which says thatthe variety (possible output states) of a control system must at least equalthe variety of its environment. A true statement, but you can't use it todesign a system having the requisite variety. It's an after-the-factdescription.
> Note that in the (3D) Albus (perhaps giving him too much credit)architecture, nodes on the same level interact directly.
While this idea could be added to the HPCT model, we haven't tried it,because so far we haven't done any experiments complicated enough to call forit. Some day, maybe.
> BTW - have you tried to cause the object being tracked in yoursimulation to change velocity in 3 space randomly? I suspect the little manwould have trouble pointing at the target.
I didn't mention the visual control level in the Little Man, which usesbinocular vision to generate x, y, and z signals for the visually-observedpositions of the fingertip and a movable target. When you get the demo disk,you'll see. The Little Man tracks an arbitrarily moving target with itsfingertip in real time. Version 1 does this better than Version 2 (the one withfull dynamics, muscle model, reflex loops) because in version 2 the dynamics ofthe visual system are poorly implemented in relation to the arm dynamics. Stillno inverse kinematics or dynamics, however, and they won't be needed even whenI clean up the top level of the model (some day real soon now).
> You might find that an environmental model incorporating the physicsof motion was needed.
Well, such a model exists in version 2, converting the torque outputs ofthe control system model into motions of the physical arm, with all theinteractions, Coriolis forces, etc. . The model of the control system, however,does no such calculations. The program only does them to show how the signalscause the physical arm to move, turning the control system's outputs intoperceptual inputs of joint angles. Only forward dynamics and kinematics areused to simulate a real arm.
> What happens if we make the little man observe in a mirror?
It would be necessary to switch a sign-inverter into the pathway carryingthe x position signal (or the error signal) from each eye. Otherwise positivefeedback would occur in the x direction. That would take another level ofcontrol.
> How do you correct your output signals when the error signal doesn'tdiminish as expected?
Since there's no prediction of how the error signal will diminish, there'sno expectation (in the model, that is -- in me there is a definite one!).There's a provision in the model for switching gravity on and off: it makeshardly any difference in behavior, although you can see the muscle forceschanging to compensate. The model handles disturbances in the usual way.
> How about at the level of verbal communication? I construct some setof statements to influence your behavior. This is based both upon my perceptionof your response, my understanding of language, and my predictions of yourreaction. This seems a bit beyond a simple error signal as there isconsiderable lag and indirection.
We've talked about such things now and then on the net, but haven't reachedany conclusions about how to model the situation. The big problem, of course,is that we don't understand how language works. That kind of modeling is a longway in the future.
> It does seem based upon some set of predictive world models, or if youprefer, simulations.
I like "simulations" better. Yes, this is true. It's on the agenda, for mygrandchildren (or someone's). Lots of fertile ground there for research.
Best, Bill P.
Date: Tue Jul 05, 1994 10:46 am PST
Subject: Albus waits
From Tom Bourbon [940705.0924]
Last Friday I said that over the weekend I would remove my state ofignorance and uninformedness about Albus, 1991, but it turns out I was alsoignorant and uninformed about the hours of the local university librariesduring the Independence Day weekend. Albus 1991 must wait until thisevening.
I did re-read the Albus series that ran Byte, 1979, in parallel with theseries by Bill Powers. Now I see why I never took seriously anyone's suggestionto read Albus, in more recent years. There are a few points in common betweenAlbus and Powers -- some of the _topics_ they discuss and some of the _words_they use (including hierarchical perception and control [but of what and bywhat?], feedback [but of what, to what?], the analog nature of synaptic eventsin nervous systems, the idea that _all_ behavior [no matter how "cognitive" itmay seem] is motor behavior, and so on) -- but the differences are glaring.I'll summarize my interpretation of those differences after I've read Albus,1991; perhaps he made some radical changes in his "model" since 1979 although Idoubt that is the case.
Later, Tom
Date: Tue Jul 05, 1994 10:52 am PST
Subject: Re: PCT models
[Paul George 940705 09:30 EDT] [Bill Powers (940702.0400 MDT)]
I am not really a control system engineer, though I have some experiencewith programming robot controllers. I only recently joined a control systemscompany, and the people I am working with are kind of on the leading edge ofcontrol theory. I may thus have a biased view of how control engineers think.By profession I am an engineering process and methods specialist (i.e humanprocesses).
> As I read this sentence, it says that the control system movessomething to particular external spatial coordinates, in an objectivecoordinate system.
There may be motion in the real universe, but the control system can neverKNOW. It has an internal coordinate system it uses as a reference grid. Itorients things it perceives in its environment in that system. But the systemonly detects the 'outside world' through it's senses.
>> There also exist reference variables to which this object 'state'is compared (albeit by changing formulae and sets of 'controlledvariables').
> I understand comparing one signal against another, but how is such acomparison done by "changing formulas and sets of controlled variables?" Didyou really mean that the _comparison_ is done by these means?
Yes. In the higher levels of distributed control systems what constitutesan error is a function of history and sets of control variables. Which inputsare important changes, as do the reference values. The system is usingsomething like statistical process control(as you note later) to determinewhether the environment appears to be changing in the desired directions, or isgoing out of control (a fuzzy concept). The system is trying to predict thefuture and take pro-active action.
Re actual application to a robot. I vaguely recall that there were some toyrobot arms that are programmable. There are also some computer games where oneprograms a robot's behavior to interact with other robots. These would be a lotcheaper that an industrial quality robot. Sorry I can't give more solidreferences. BTW I don't know if you have seen it, but some research hassuccessfully used chaos formulas to allow a multi-legged robot insect to walkwithout directly controlling the individual leg motions. The old joke about thecentipede tripping when it paid attention to its feet appears to have a grainof truth.
> However, it's not clear in what sense your "mirrored object" reflectsthe real world entity. ..... We have toyed with the idea of actual analogmodels of the external world being used by the control system, largely to fillin missing information, but so far have not come across a case that requiresthis (not to say that such cases will not be found)
Objects were being used in the Object Oriented Analysis & Design sense.By convention we try to have objects in the control system software thatcorrespond to the real world entities. They 'own' both the reference variables(in our terms instance variables) and the functions (methods) which may be usedto influence the object. The object interacts with the available sensory datato 'set' the instance variables, often algorithmically (we can rarely directlymeasure what we are interested in). We must 'reverse engineer' the realentities' state based upon a chain of indirect information. A thermocoupleproduces a voltage, which is converted into a digital 'count'. This is input tothe object and converted to a temperature value. However, often we will averagethe results of several sensors to infer the temperature of the material in thevessel. Further, we are usually interested in the temperature trend, whichinvolves assumptions about the heat capacity of the material and the energy ofthe heaters.
> I do not consider that the basic sensing process is an A-D conversion:it is a conversion from stimulus amplitude to impulse frequency, which is stillan analog measure.
Ack. A to D conversion is an artifact of computers and electronics. I wouldbe surprised to find it (as a general mechanism) in living systems. My weakunderstanding of nerve dendrite interaction leads me to believe that thesignals are fairly complex.
Note that my descriptions are of what plant control systems use. It may ormay not apply to living systems, though it hints that data and modelingmechanisms _may_ be an efficient architecture at higher levels ofcontrol(particularly with multiple entity systems - e.g. social). OTOH,biological analogies are becoming fashionable in the control industry.
>> The key idea is that granularity of data (input & outputsignals, reference & control variables) and time horizon increase as you goup the hierarchy.
> Well, that may be true, but it's not a very useful concept for makingmodels. It's sort of like Ashby's "Law of Requisite Variety," which says thatthe variety (possible output states) of a control system must at least equalthe variety of its environment.
It is useful, and basic. The idea is that the sensing rate and the'reaction time' slows as you go up the hierarchy. Higher levels have a wider'span of control' (Management Science(?) concept), care only about 'summaryinformation' and have a longer time horizon. Higher levels can sample thecontrol variables of lower levels (among other things).
More directly to the point, one of the key design problems of safetycritical systems is determining the complete state space of the environment,and insuring all unsafe states are detected. Fortunately, most states andevents are of the 'don't care' variety. The key concept of a model(and acontrol system) is that it handles the 'important' behavior and variables ofthe real world system.
>> What happens if we make the little man observe in a mirror?
> It would be necessary to switch a sign-inverter into the pathwaycarrying the x position signal (or the error signal) from each eye. Otherwisepositive feedback would occur in the x direction. That would take another levelof control.
Precisely my point. How does the 'little man' tell that a mirror has beeninserted? How does he (or more fairly a sophisticated control system like anorganism) detect the positive feedback and correct the comparison algorithm(Ilook at another level of control as a modification of the control logic)? As Ihave heard it tossed around, the concept of 'reorganization' of the controlsystem seems a bit like a Deus ex machina. There needs to be some mechanism todetect that the system is out of control (or trending that way) and that thecontrol system needs to adapt. Something like a deductive mechanism wouldappear to be needed.
> It's on the agenda, for my grandchildren (or someone's). Lots offertile ground there for research.
What?? You _can't_ give us all the knowledge of the world while standing onone leg? How disappointing :-)
(From Tom Bourbon [940701.1601])
> I asked you how you go about assessing models and theories -- whichcriteria and procedures do you use before you decide a theory or model is"good" or that it "works?"
I thought I had answered. We create a model when we can't handle thecomplexity of the real world, or if we need answers in faster than real time.The model should exhibit the same behavior as the modeled system or allow itscomputation, in terms of the factors of interest (aye, there's the rub). Amodel is good IF it is useful for the intended purpose. It need not beaccurate, in the sense of using the exact mechanisms as the real world, unlessthat is its intent. It almost by definition cannot be complete in the sense ofcapturing all the real system's complexities, again unless that is itsintent.
Models can be for several purposes. One variety is for understanding, as inan analogy, Einstienian thought experiment, or scientific theory. Such modelsallow us to reason about how the real system should react, and deal withcurrently 'unexplained' effects. Another type mimics a portion of the realsystem in order to allow analysis or temporal prediction, as in asimulation.
The model is good if it allows us to come to the 'same' conclusions as thereal world system, or allows us to obtain our desired results. A theory shouldexplain all observed behavior of interest. A simulation's outputs should berelated in a known (deterministic) way to the real system's outputs, within acertain range of inputs and outputs.
Alchemy works to a certain extent, though quantum mechanics works betterfor explaining how chemical compounds will behave. Natural selection helpsexplain ecological adaptation and speciation. Fluid dynamics and aircraftsimulations allow us to analyze airframes (and train pilots) without litteringthe countryside. The validation of the model is what happens with the realplane.
Albus' model is good if it allows us (my company) to build a more robustand easy to manage control architecture. It is partially based upon observationof natural and artificial systems. Nature has devised mechanisms that work, andif we can incorporate validated biological control architectures to avoidre-inventing the wheel. Whether it in fact is the general architecture forintelligence, remains to be seen.
With PCT, the question is the intent of the model. Is it to be a 'universalfield theory' for biological systems, a biophysical model of a nervous system,or some other purpose? The simulations you have built appear to show that themodel provides a mechanism which mimics living systems at the reactive level.It appears to be based upon what we observe (IMIO) at the neural level. It maycome to pass that it can be used to predict or understand the behavior ofsocial groups or cognition. However, because the model behaves the same asnature, it does not necessarily follow that the PCT mechanism is that which isused by nature at all levels. It might however be a very powerful technique forengineering PLC level industrial controls.
Perhaps PCT is the mechanism used at the biological level, and Albus' isthat used at the cognitive or inter-personal level. Perhaps over the next fewgenerations we will find out. It would surprise me if the exact same mechanismwas used at all levels of control. Genetic drift and thermodynamics would seemto give examples that different mechanisms often provide similarfunctions.
Date: Tue Jul 05, 1994 7:49 pm PST
Subject: Re: PCT models
From Tom Bourbon [940705.1442] >[Paul George 940705 09:30EDT]
. . .
>(From Tom Bourbon [940701.1601])
>> I asked you how you go about assessing models and theories --which criteria and procedures do you use before you decide a theory or model is"good" or that it "works?"
> I thought I had answered.
Perhaps I didn't state my questions as clearly as I thought, either time. Iasked how you assess models and theories _in general_ and then I asked if youwould tell us specifically how you applied your preferred criteria to anassessment of the model of Albus. I am simply trying to gain an understandingof how and why you believe an examination of the Albus model might be of valueto PCT modelers.
> We create a model when we can't handle the complexity of the realworld, or if we need answers in faster than real time.
This sentence heightens my interest in learning specifically how youassessed the Albus model and found it good. You have described uses of modelsthat are somewhat different from the way we use them in PCT. In "Models andtheir worlds," Bill Powers and I talked a little about the differences betweenuses such as you described here and our use of models in PCT. We use them in away that is common in engineering -- to test the functionality of what we thinkis an explanation for an instance of control -- we devise a PCT model of thesystem we think is producing control in a particular situation, than run themodel to determine if our ideas are "legitimate," or if they are allwet.
> The model should exhibit the same behavior as the modeled system orallow its computation, in terms of the factors of interest (aye, there's therub).
Here, we agree, but I wonder how you test for the behavior of an Albusmodel? I ask this in earnest; I cannot imagine (my failure) how one would setup a working model of Albus's conjectures, then run it in simulations thatyield bouts of continuous quantitative control behavior.
> A model is good IF it is useful for the intended purpose. It need notbe accurate, in the sense of using the exact mechanisms as the real world,unless that is its intent. It almost by definition cannot be complete in thesense of capturing all the real system's complexities, again unless that is itsintent.
If it is offered as a model for systems that produce control, then themodel should recreate the control behavior of the modeled system(s). It shoulddo that quantitatively and accurately, moment by moment. Shouldn't it? (If not,then we are not engaged in the same kinds of modeling activity.) If so, howdoes one go about setting up and testing an explicitly Albus-like model forcontrol phenomena? I am being very concrete about this question, which mayexplain why I didn't recognize the fact that you thought you had answeredme.
> Models can be for several purposes. One variety is for understanding,as in an analogy, Einstienian thought experiment, or scientific theory. Suchmodels allow us to reason about how the real system should react, and deal withcurrently 'unexplained' effects.] Another type mimics a portion of the realsystem in order to allow analysis or temporal prediction, as in asimulation.
We may be talking from different sides of that distinction. I believe Albussaid he offered his model as a speculation that would inspire empirical work byother people; we are trying to do the empirical work, using a differentapproach to building and testing a model that I see in Albus's work.
> The model is good if it allows us to come to the 'same' conclusions asthe real world system, or allows us to obtain our desired results. A theoryshould explain all observed behavior of interest. A simulation's outputs shouldbe related in a known (deterministic) way to the real system's outputs, withina certain range of inputs and outputs.
Fine. And which criteria would one use to determine how well a simulation'sresults are related to the real system's outputs? (Actually, in PCT we aren'tjust interested in the system's outputs; we are more interested in the_consequences_ of the outputs of either a model or the living system itrepresents.) We typically look for very high correlations between variablesassociated with the behavior of the model and of a person, or very low rmserror between predicted and actual results, or some other strict quantitativeagreement between predicted and actual results.
I'll defer on answering the remainder of your reply to me until after Iread Albus, 1991. I finally have a copy.
Later, Tom
Date: Tue Jul 05, 1994 7:52 pm PST
Subject: Re: PCT Models
[From Rick Marken (940705.1340)] Paul George (940705 09:30 EDT)
> With PCT, the question is the intent of the model.
This is the question, indeed. The intent of the PCT model is to explain thephenomenon of control as it is manifested in the behavior of living systems.PCT is not a newer or snazzier version of control theory; it is just plain oldcontrol theory. What distinguishes PCT from all other models of behavior (suchas Albus') is HOW it maps control theory to behavior. PCT begins with theobservation of a phenomenon -- control. "Control" refers to the fact thatorganisms reliably produce consistent results in an inconsistent environment.People walk, talk, go to work, plant tress, cut down trees, eat, go to bed, etcetc -- and they do these things over and over again, always under slightly (andsometimes profoundly) different circumstances.
Powers recognized the existence of the phenomenon of control and saw thatit was what had always been called "purposeful behavior". Control occursbecause organisms are able to vary their effects on the world, as necessary, inorder to produce the results they intended, all the while preciselycounteracting unpredictable (and typically undetectable) disturbances to theseresults. Powers saw that no existing psychological theory explained howorganisms could possibly do this -- ie. how they could control. Even controltheory, as it had been applied in the life sciences, didn't explain thephenomenon of control; control theory (before PCT) had been used to describethe relationship between inputs; non-PCT control theory is really just S-Rtheory in the Laplace domain ;-).
Powers' figured out how to correctly map control theory to actual controlbehavior. When you apply control theory correctly to the phenomenon of controlyou find that living control systems control perceptual input variables andthat their outputs depend largely disturbances and variations in the feedbackfunction relating output to input -- factors that are independent of the livingsystem itself.
It is impossible (I think) to appreciate the extraordinary significance ofPCT without a thorough understanding of the phenomenon that it explains --control. This phenomenon is so ubiquitous that it is almost invisible. But youcan learn to see it -- sometimes by just looking at it and other times byactively testing for control via the introduction of disturbances tohypothetical controlled variables.
The life sciences, and the behavioral sciences in particular, have missedthe phenomenon of control completely; all behavioral science research andmodelling (to the extent that it occurs) is based on the assumption that livingsystems respond to stimuli or emit behavioral outputs; in fact they control.Thus PCT, which explains a phenomenon that doesn't even exist, from the pointof view of conventional behavioral science, has been almost completelyignored.
PCT is not a sexy new model, like chaos, fuzzy logic, dynamical systems,etc, because the main goal of PCT is not to develop a complex theory but tounderstand a real phenomenon (you can probably guess what that phenomenon is bynow). There are some fairly rich PCT models -- such as Bill Powers'hierarchical arm pointing system and my three level spreadsheet model with sixsystems controlling a different type of variable at each level -- but these arejust demonstrations of what can be done with the hierarchical model. The mostimportant work in PCT (I think) involves demonstrations of real controlphenomena.
PCT is unique (and revolutionary -- unfortunately) not because it is a newtheory but because it deals with a phenomenon that is not dealt with explicitlyby any other theory -- and because it properly applies the existing engineeringmodel of control to the controlling done by living systems.
Caveat emptor: Just because a theory has the word "control" in its name ora theorist talks about "control of this" and "control of that" does not --repeat, DOES NOT -- mean that the theory or theorist is dealing with thephenomenon of control. In fact, I don't believe that there is ANY theorybesides PCT that is an explicit attempt to explain control (but I would love tobe surprised).
So accept no substitutes; if understanding control is your game then PCT isthe name :-)
Best Rick
Date: Wed Jul 06, 1994 7:52 am PST
Subject: Re: PCT models
[Paul George 940706 11:00 EDT] >Tom Bourbon [940705.1442]
> I am simply trying to gain an understanding of how and why you believean examination of the Albus model might be of value to PCT modelers.
It may not be of immediate use to PCT modelers (focusing on PCT control inthe small), but It might be to PCT theorists focusing on higher levels of thehierarchy.
I found the overall architecture appealing aesthetically, and applied theframework to both control systems and organizations in thought experiments. Itseemed to fit quite well as a technique that might work to rationalize bothtypes of systems (note that my focus is on organizational and methodologicaltechnology for development projects). Its separation of sensor/actuatorjudgement, and a world model (perception/simulation) using autonomous objectlike structures seemed a good differentiation approach (division of labor &recognition of separable error sources). Organisms have different cell typesand organs. Organizations have roles and departments.
Having now read Dag's data packet and Power's posts over the last few days,I think there are strong correspondences between Albus' architecture & PCT,particularly at the individual cognitive and organizational levels. Inparticular, both use planning, History/memory, goal seeking,configuration/influence and data passing of between lower and higher levelnodes, an effective model or simulation of the putative 'real world'. If youput the PCT concepts of comparison in the value judgement node, and theconcepts of controlled perception in the transducer and world model nodes, PCTadds 'mechanism' to the Albus architecture. The exact nature of the signalstransmitted between nodes is almost in the noise range from an architecturalstandpoint.
I find it interesting that you appear to come from the standpoint that youshould by default _not_ consider another's work, particularly from a 'rival'school. My approach is to examine any theory to see if it contains anythinguseful that might provide insight into others, if nothing else than anotherpoint of view. There is a tendency in the academic and methodological (a.k.a.consulting) worlds to focus on details for the purpose of differentiation.Brand names and ideological purity are the basis of recognition and reputation.One must deny any validity to another's approach in order to make one's ownmore prominent. "I cannot be RIGHT unless everybody else is WRONG" (orinferior, or at least different and incomparable). "Ignore them, they are notof the faithful".
> We use them in a way that is common in engineering -- to test thefunctionality of what we think is an explanation for an instance ofcontrol
Ah. This may be the basis of confusion. I was using model in the generalsense of a theory - PCT _is_ a model of behavior. Executable models may be usedto test mental models, or at least determine their feasibility. I wouldnormally call this an experiment or simulation. It is however not the _only_method of testing a theory. The experiments you have done (sorry still haven'tgotten access to your book)
> Here, we agree, but I wonder how you test for the behavior of an Albusmodel? I ask this in earnest; I cannot imagine (my failure) how one would setup a working model of Albus's conjectures, then run it in simulations thatyield bouts of continuous quantitative control behavior
Easily. Albus proposes an organizational framework, not a model in the PCTusage. We observe that control and information systems have been built withinput and output handlers separate from the 'main' program logic. Similarlyapplications often separate the knowledge base and algorithmic processing.Distributes control systems (industrial) are hierarchical and show the timehorizon behavior. Albus attempts to suggest a general framework abstracted fromspecific instances. Creating a control application using the 'object classes'identified in Albus' paper is no great technical challenge (we're doing it).The test is a system or organizational architecture that is simple and works.OTOH I may be unclear on your goal. " bouts of continuous quantitative controlbehavior" is not terribly meaningful to me, except at a limited scope - that ofneurophysiology or device and PLC level automatic control.
> ....And which criteria would one use to determine how well asimulation's results are related to the real system's outputs?
If you want to test whether the model works for _human_ behavior, than youmerely ;-} have to program the various object methods and attributes tocorrespond to your understanding of the specific variables and comparators, andsensory/communications distortion functions. Then you can see if it works likethe real thing, as you later describe. But, you are testing your control logicand reference variables & values, _not_ the architecture.
Date: Wed Jul 06, 1994 12:00 pm PST
Subject: Re: PCT models
From Tom Bourbon [940706.1227]
Thanks, Paul, for you thoughtful reply to my previous post. I have readAlbus, 1991 (and re-read Albus, 1979) and I am still trying to decide how tosummarize that material -- my own notes on the articles add up to over 18pages, so I _must_ try to identify the key points. Your post is helping me zeroin on them.
>[Paul George 940706 11:00 EDT]
>>From Tom Bourbon [940705.1442]
>> I am simply trying to gain an understanding of how and why youbelieve an examination of the Albus model might be of value to PCTmodelers.
> It may not be of immediate use to PCT modelers (focusing on PCTcontrol in the small), but It might be to PCT theorists focusing on higherlevels of the hierarchy.
Perhaps, but I'm not sure that will prove to be the case. When I decide howbest to summarize Albus, I'll try to identify my reasons for that feeling orbelief.
> I found the overall architecture appealing aesthetically, and appliedthe framework to both control systems and organizations in thought experiments.It seemed to fit quite well as a technique that might work to rationalize bothtypes of systems (note that my focus is on organizational and methodologicaltechnology for development projects).
I believe this is one of the primary differences in our perspectives onAlbus. Understandably, you are interested in designing organizations andsystems that achieve your intended ends -- or those of your clients. Some ofthe uses for which you design systems would overwhelm any single person actingas a control system (a topic addressed nicely by Bill Powers (940706.0630 MDT).In that role, you can (and should) use any trick you can either dream up, orlocate already in use somewhere. In his "model," Albus has done the same thing;he has put together an overwhelming array of bits and pieces from robotics,traditional control engineering, neurophysiology, psychology, popularliterature and other sources, all in a way that would understandably haveaesthetic appeal to someone who is looking for any conceivably useful tools heor she can find. I have no problem either with what Albus has done, or withanyone who uses his "model" as a source of ideas and techniques that will helpin designing and building systems for clients.
I have no problem, that is, _unless_ Albus's "model" is presented as amodel for living things. I am tempted to say that it is not such a model, but Iwill simply say that if it is intended as a model of living things, then it isnot the same kind of model as the one in PCT. More on that idea when I puttogether my summary of Albus.
> Its separation of sensor/actuator judgement, and a world model(perception/simulation) using autonomous object like structures seemed a gooddifferentiation approach (division of labor & recognition of separableerror sources). Organisms have different cell types and organs. Organizationshave roles and departments.
I agree that organisms have different kinds of cells and organs. I alsoagree that organizations have roles and departments. Let's call those twostatements "facts." I do not see that the one fact has any significance for theother. My inability to see such significance does not imply a rejection by meof the idea that you, or anyone else, might see it. It is up to me (rather, itis up to PCT modelers) to demonstrate that our way of modeling living thingscan succeed without relying on the relationship of one such fact to theother.
> Having now read Dag's data packet and Power's posts over the last fewdays, I think there are strong correspondences between Albus' architecture& PCT, particularly at the individual cognitive and organizational levels.In particular, both use planning, History/memory, goal seeking,configuration/influence and data passing of between lower and higher levelnodes, an effective model or simulation of the putative 'real world'. If youput the PCT concepts of comparison in the value judgement node, and theconcepts of controlled perception in the transducer and world model nodes, PCTadds 'mechanism' to the Albus architecture. The exact nature of the signalstransmitted between nodes is almost in the noise range from an architecturalstandpoint.
In some ways, I see what you are saying and I can agree. But in other ways-- major ways -- I believe there are significant differences between the twoarchitectures. As you say, PCT includes a specific mechanism (the inputfunction, comparator function, and output function, with their accompanyingsignals -- perceptual signal, reference signal and error signal), while Albusdoes not. But I do not believe one can so easily map PCT functions and signalsonto the Albus architecture. For one thing, all of the components in the PCTmodel are "dumb," in that each of them (except perhaps the input functions)performs a relatively simple operation; in contrast, the elements and modulesin Albus's "model" are often very "smart" indeed, with each one of themperforming many different functions in many different ways. More on this in mysummary.
> I find it interesting that you appear to come from the standpoint thatyou should by default _not_ consider another's work, particularly from a'rival' school.
That is not at all the case. I must have overspoken in some of my previousposts. I had read Albus's earlier work and concluded that it would not help usin our attempts to model living systems, _as much as possible_ in their ownterms and from their own perspectives. My decision was not made out of anydefault rejection of work by other people; I believed Albus was simply talkingabout a different kind of game -- one I can fully accept and respect, up to thelimit I described above. Now that I have read Albus, 1991, my assessment hasnot changed.
> My approach is to examine any theory to see if it contains anythinguseful that might provide insight into others, if nothing else than anotherpoint of view.
I (and I believe Bill Powers and Rick Marken) did just that with Albus, andwe independently came away with the idea that his work did not provide insightswe could use. We could be wrong.
> There is a tendency in the academic and methodological (a.k.a.consulting) worlds to focus on details for the purpose ofdifferentiation.
True. But there are also real differences between various theories andmodels. In science, people who observe such differences are obligated to makethem known.
> Brand names and ideological purity are the basis of recognition andreputation. One must deny any validity to another's approach in order to makeone's own more prominent. "I cannot be RIGHT unless everybody else is WRONG"(or inferior, or at least different and incomparable). "Ignore them, they arenot of the faithful".
That may be true (to some degree or other) for people who play the game ofBig Science. I do not knowingly or intentionally play that game. If PCTmodelers cared about such things, we all would have opted for other "brands"long ago! ;-))
We _do_ say things like, "Here is an example of control and here is ourmodel for the system that produces control; you are invited to show us anyother model that you believe performs as well as or better than ours." As ourreward, we have had editors and reviewers refer to that approach as "cute" anda "clever ploy."
>> We use them in a way that is common in engineering -- to test thefunctionality of what we think is an explanation for an instance ofcontrol
> Ah. This may be the basis of confusion.
I had begun to suspect that this might be a point where we were usingsimilar words and talking past one another.
> I was using model in the general sense of a theory - PCT _is_ a modelof behavior.
I thought so.
> Executable models may be used to test mental models, or at leastdetermine their feasibility.
Yes.
> I would normally call this an experiment or simulation.
No problem; so would we, if editors and reviewers in behavioral sciencewould allow it.
> It is however not the _only_ method of testing a theory.
Certainly it is not, but it is a rigorous one -- a strict one -- and weprefer to apply just such a method when we test our ideas about models forliving control systems.
>> Here, we agree, but I wonder how you test for the behavior of anAlbus model? I ask this in earnest; I cannot imagine (my failure) how one wouldset up a working model of Albus's conjectures, then run it in simulations thatyield bouts of continuous quantitative control behavior
> Easily. Albus proposes an organizational framework, not a model in thePCT usage. We observe that control and information systems have been built withinput and output handlers separate from the 'main' program logic. Similarlyapplications often separate the knowledge base and algorithmic processing.Distributes control systems (industrial) are hierarchical and show the timehorizon behavior. Albus attempts to suggest a general framework abstracted fromspecific instances. Creating a control application using the 'object classes'identified in Albus' paper is no great technical challenge (we're doing it).The test is a system or organizational architecture that is simple and works.OTOH I may be unclear on your goal. " bouts of continuous quantitative controlbehavior" is not terribly meaningful to me, except at a limited scope - that ofneurophysiology or device and PLC level automatic control.
So the criteria by which you judge success or failure are different fromthose we elect to apply when we test our ideas. I have no problem with that.Incidentally, for those who have not read Albus's papers, you have just given anice glimpse of how he describes a system and its workings; those who arefamiliar with PCT may begin to see why I am wrestling with how to characterizethe similarities and differences between Albus and Powers! :-)
Later, Tom
Date: Wed Jul 06, 1994 5:24 pm PST
Subject: Re: PCT models and simulations
[From Bill Powers (940706.1210 MDT)] Paul George (940706.1100 EDT)
RE: Albus' model.
> I found the overall architecture appealing aesthetically, and appliedthe framework to both control systems and organizations in thought experiments.It seemed to fit quite well as a technique that might work to rationalize bothtypes of systems...
Unless you are a very competent modeler, dabbling in thought experimentscan be the basis of delusions more easily than truth. The problem with doingthought-experiments on aesthetically-pleasing patterns of system design is thatyou _think_ you know what a real system with that kind of design will do, butyou can easily be completely wrong. The fact that Albus claims that a systembuild according to his design would behave in a certain way is no indicationthat such a system could even be built, or if built that it would actually doanything like what he thinks it would do.
At one point, some psychologists heard about positive feedback, andthinking that anything "positive" must be good, they starting musing about theeffects of "deviation-amplifying responses" and such things, coming up with allsorts of fanciful tales about the advantages of positive feedback. Perhaps byamplifying deviations, a system could increase its variety and so better matchthe variety of the environment. They imagined that such systems would haveuseful properties, but they left out one critical step: actually analyzing,building, or simulating a positive-feedback system to see how it would reallybehave. Being, in fact, incompetent modelers, they didn't realize that anactual system organized in the way they proposed would have properties theyhadn't anticipated, and would in fact blow itself up in a few seconds, or elsedo nothing at all of any interest.
If there's one thing any competent modeler knows, it's that any complexsystem design, unless it's thoroughly familiar and completely analyzed, isgoing to have important modes of operation that are different from what onethought they would be, and even more important modes of operation that aretotal surprises. Anybody can sit down and draw a fancy system diagram with allthe parts labelled with function-names, with nice-looking symmetries andimpressive patterns of repetition. Furthermore, anyone can claim that such asystem would behave in certain ways. That sort of stuff is a dime a dozen;anyone with a fertile imagination can do it. But if that's all that is done,the result is just a fairy-tale.
The whole trick in successful modeling is to PROVE that a system designedin a certain way will actually behave as you imagine it will behave. There areonly three ways I know of to do this. One is to do a complete mathematicalanalysis of the components and their connections and find analytical solutionsof the system of equations. That is almost never possible for any system ofeven moderate complexity. The second way is to build a physical system usingcomponents having the specified properties and turn it on. That is usually tooexpensive an approach, especially considering that the system is going torequire many revisions before it does anything like what you have imagined. Andthe third way is to design a simulation in which the properties of eachcomponent are accurately portrayed as elements in a working computationalmodel, and run the simulation. Most of the time, that's the only practical wayto test an idea.
The reason that Tom Bourbon (and Rick Marken and I) is skeptical aboutmodels that haven't been tested in simulation is that the human brain is verygood at guessing wrong about what systems will do. We let what we WANT thesystem to do persuade us that our design will actually do what we want. If thesystem contains any reasonable amount of complexity, we will usually find thatwe have designed parasitic loops into it, or subtle self-contradictions, orhave required some variable to do something physically impossible such as goingfrom positive to negative values without crossing zero, or have built inrequirements that assume a physical component capable of behaving with 18decimal places of accuracy, forever. A competent modeler anticipates suchproblems and tries to take them into account. The neophyte just sails right in,drawing diagrams and claiming that they explain something.
When you sit down to produce an actual simulation, the most common thingthat happens is that you come across something you have assumed to happen thatthe system you have actually designed can't do -- or that you have completelyfailed to provide for in the model. One of my arguments with Martin Taylorabout the "alerting system" is based on this problem. It's easy to propose asystem that will alert other systems to a problem. But when you sit down tosimulate such a system, you discover all the machinery you were unknowinglyassuming and that is essential to create the situation you had in mind.Providing this machinery very often turns out to be a far larger problem thanthe original design. And you can't get a simulation to run properly unless youinclude in it EVERYTHING it needs in order to run. You can wave your arms allyou like; the system will still do only what you gave it the capability ofdoing. Also, like a computer program, it will do exactly what you designed itto do, not what you wanted it to do.
Rick Marken and Tom Bourbon and I have all tried to make system designsproposed by psychologists work. And I mean we have really tried. We have evensupplied missing properties in our attempt to make such simulations produce thekind of behavior that the proponents of psychological theories claim theirsystems would produce. And despite the fact that we are fairly competentmodelers, we have failed. Just try designing a simulation of anoperant-conditioning experiment according to traditional explanations, one thatwill reproduce what is actually observed. You immediately discover that thereis no specification for the effect of reinforcement on behavior that can beturned into a model. When you try to supply one, you find that translations ofthe usual verbal descriptions result in models that either don't do anything atall or settle into modes of behavior that are never observed in real systems.We have publicly challenged psychologists to supply their own designs forsimulations, and have even offered to do the programming work ourselves if theydon't know how. These challenges have, of course, disappeared into an echoingsilence, because the psychologists involved don't have the vaguest idea of whatwe are talking about. They think that if they _say_ that reinforcement worksthe way they assume, such a system _will_ work as they say it does. The idea ofactually testing these assertions has never crossed their minds.
One reason that PCT models tend to be simple is that we can't prove thatany more complex models would actually behave as we imagine they would. All themodels we talk about have been proven to work exactly as claimed, as well asmatching real behavior with acceptable accuracy. The idea of jumping intoextremely complex behavioral problems while we still can't simulate simplerbehaviors is just ridiculous. We're building upward from a solid base. All theguesses as to what we will find as models of higher levels of control are justthat, empty guesses. Whatever we guess right now is almost certain to be provenwrong; why waste time in airy-fairy-land? Why bluff when we can stick with whatwe know about?
Best, Bill P.
Date: Thu Jul 07, 1994 6:10 am PST
Subject: Albus articles
From Tom Bourbon [940707.0813]
Paul George is the most recent person to suggest that PCT modelers mightfind useful ideas in the writings of James Albus. (Some people have alsosuggested that the model in PCT is a special, limited, case of a more generalmodel of control developed by Albus.) For some years now, Albus has writtenabout robots and artificial systems. He uses a particular interpretation ofideas from psychology and neurophysiology and combines them with ideas fromwork on artificial systems, with the intention of suggesting an "outline" or"framework" to be used in developing a general theory or model ofbehavior.
After seeing Paul's suggestion, I re-read four articles by Albus that ranin _Byte_ magazine, June-September 1979, concurrently with a series on controlsystem theory by Bill Powers. I also read the more recent article suggested byPaul:
J. S. Albus (1991). Outline for a theory of intelligence. IEEE Transactionsfor Systems, Man, and Cybernetics, 21(3), 473-509.
I took more than 18 pages of notes on the articles. I cannot possiblysummarize that much material here. All I present here is a brief description ofa few of my conclusions. I use several quotes from Albus, thereby avoiding theneed to present my own interpretations of him, which might be biased. I urgeanyone who wants to challenge or check my conclusions to read the same articlesand to perform the modeling I suggest at the end of this post.
The Elemental Albus Architecture:
Albus proposes the following architecture for his theory. I believe thediagram and a brief summary of his descriptions of the various elements in itclarify many differences between his theory and PCT.
____
Perceived Situations | | Plan Evaluations
,--------------------------->| VJ|-----------------------------,
| |____| |
| Situation | | |
| Evaluation | /|\ Plan Results |
| | | |
| \|/ |\|/
__|_ Updates |__| Plans_|__
| |------------------------->| |<--------------------------||
| SP | | WM | | BG|
| |<-------------------------| |-------------------------->||
|____| Predicted Input |____| States|____|
| |
/|\ |
| Commanded Actions |
| Observed Input |
|_____ ______\|/
| | | |
|Sensor| |Actuator|
|______| |________|
/|\ |
| ___________ |
| Events | | Actions |
'----------------------|Environment|<--------------------'
|___________|
The system portion of this busy looking elemental loop repeats in ahierarchical manner, much as the elemental loop in PCT repeats in ahierarchy.
DEFINITIONS AND DESCRIPTIONS OF ELEMENTS IN ALBUS'S ARCHITECTURE
Albus's sensors and actuators are essentially like the input and outputfunctions in PCT. Very little else about his architecture is like anything inPCT. All of his other modules or elements are "intelligent" -- loaded withfunctions, powers and abilities unlike anything in the simple elements of a PCTmodel. I quote from Albus. Pay close attention to the various "systemelements," and to the bewildering array of abilities, roles, powers, andfunctions he attributes to those elements and systems.
SP = "Sensory Processing. Perception takes place in a sensory processingsystem element that compares sensory observations with expectations generatedby an internal world model. Sensory processing algorithms integratesimilarities and differences between observations and expectations over timeand space so as to detect events and recognize features, objects, andrelationships in the world." p. 476.
"Perception is the establishment and maintenance of correspondence betweenthe internal world model and the external real world." p. 493.
WM = "World Model. The world model is the intelligent system's bestestimate of the state of the world. The world model includes a database ofknowledge about the world, plus a database management system that stores andretrieves information. The world model also contains a simulation capabilitythat generates expectations and predictions. The world model thus can provideanswers to requests for information about the present, past, and future statesof the world. The world model provides this information service to the behaviorgeneration system element, so that it can make intelligent plans and behavioralchoices, to the sensory processing system element, in order for it to performcorrelation, model matching, and model based recognition of states, objects,and events, and to the value judgment system element, in order for it tocompute values such as cost, benefit, risk, uncertainty, importance,attractiveness, etc. The world model is kept up-to-date by the sensoryprocessing system element. p. 476.
VJ = "Value Judgment. The value judgment system element determines what isgood and bad, rewarding and punishing, important and trivial, certain andimprobable. The value judgment system evaluates both the observed state of theworld and the predicted results of hypothesized plans. It compares costs,risks, and benefits, both of observed situations and of planned activities. Itcomputes the probability of correctness and assigns believability anduncertainty parameters to state variables. It also assigns attractiveness, orrepulsiveness to objects, events, regions of space, and other creatures. Thevalue judgment system thus provides the basis for making decisions -- forchoosing one action as opposed to another, or for pursuing one object andfleeing from another." pp. 476, 477.
BG = "Behavior Generation. Behavior results from a behavior generatingsystem element that selects goals, and plans and executes tasks. Tasks arerecursively decomposed into subtasks, and subtasks are sequenced so as toachieve goals. Goals are selected and plans generated by a looping interactionbetween behavior generation, world modeling, and value judgment elements. Thebehavior generating system hypothesizes plans, the world model predicts theresults of those plans, and the value judgment element evaluates those results.The behavior generating system then selects the plans with the highestevaluation for execution. The behavior generating system element also monitorsthe execution of plans, and modifies existing plans whenever the situationrequires." p. 477.
Albus says some of these "system elements" have subsystem elements -- asmany as three or four of them. At all hierarchical levels of the SP and WMelements, information is represented in both iconic and symbolic forms. "Ateach level in the control hierarchy, the difference vector between plannedcommands and observed events is an error signal, that can be used by executorsubmodules for servo feedback control modules for evaluating success orfailure." (p. 481) Differences between predicted and actual perceptions areerror signals that are used to induce changes _inside the system_, perhapssimilar to reorganization in PCT -- but I can't be sure about the degree ofsimilarity. At each level, Kalman filters make predictions, using recenthistorical data to compute parameters that are then used to make extrapolationsinto the future. (p. 480)
Emotions: Albus uses an outdated and highly questionable interpretation ofthe anatomy and neurophysiology of the "limbic system" in the brain, as asource of emotions. In the _Byte_ series, he wrote about how emotions from thelimbic system are input, along with "moderator variables," into the "will," aname he gave to the behavior-generating system element.
Reification from Observed Events: In all of the articles I read, Albusreified the observed results of behavior, then attributed many different"powers" to the reified concepts. A clear example is his definition of"intelligence," the subject of his article in 1991. He defined intelligence as,"the ability of a system to act appropriately in an uncertain environment"p.474. An observation that organisms _do_ act appropriately (intelligently) isexplained by an assumed property (trait, power, ability, etc.) called"intelligence." Reification of this sort is common in the behavioral sciencesand in behavioral neurophysiology and neuroscience. Once he assumesintelligence into existence, Albus says it (intelligence, which he described asan ability) requires further abilities -- "to sense the environment, to makedecisions, to control action" p. 474. "From the viewpoint of control theory,"says Albus, intelligence is an, ". . .integration of knowledge and feedbackinto a sensory-interactive goal-directed system that can make plans, andgenerate effective, purposeful action directed toward achieving them." From theviewpoint of psychology, he says, intelligence is a behavioral _strategy_. Oncehe brings intelligence into existence as reification from behavior, he says itis an ability, an integration, and a strategy. There is more on this theme, butI believe you get the idea.
General Conclusion:
Albus does not describe a generative model of the organization of livingsystems. Instead, he describes a highly complex flow chart of relationships(some empirical, many more conjectural), arranged in a hierarchy that includes(in conjecture) feedback loops, error signals, sensory processing withperceptions from lower levels combined to form more complex perceptions athigher levels, and a number of other features that, on first encounter, looksimilar to ideas in PCT, but he often uses those ideas differently than we doin PCT. I believe the "outline" or "framework" Albus presents may be useful tosome people, in their efforts to develop engineered systems, but it cannotserve as a generative model for control by living systems.
More to the point, as Albus described his theory in the article in 1991, Ido not believe it can be programmed as a working model of any system, living ornot. I _do_ believe someone could pick various elements from his outline, whichis like a giant buffet of elements and processes, and use some version of thatsubset to produce a working model, but that person would not be using "TheAlbus Model" in its fullness.
An Oft-repeated Invitation:
I admit that my inability to imagine how the Albus outline could be turneddirectly into a working model may be only that, my inability, but I believethere is more to it than that. Whatever I may believe about the Albus outline,I welcome simulations and demonstrations from anyone who wants to show that myconclusion is wrong. I will happily give any objecting person(s) copies of thecomputer code for the procedures Bill Powers and I used to generate the targetand disturbance functions for our paper, "Models and Their Worlds." I encouragethe objector(s) to use those procedures in the "environment" box of the Albusarchitecture, then implement a single-loop version of an Albus model, with allof the elements, connections, and processes that are shown in the figure Ireproduced above. Then the person should run the model in simulation and reportthe results so that we can compare the results from the two models when bothare applied to what many people think is a trivially simple, low-level, exampleof human performance. If the Albus and PCT models perform equivalently at thatlevel, then we can begin to talk seriously about how the two might compare asmodels of higher levels in the control hierarchy.
I am not being "cute" or "judgmental" in this offer; I simply cannot thinkof another way to compare the Albus and Powers models, as working models ofcontrol behavior.
Later, Tom
Date: Thu Jul 07, 1994 10:07 am PST
Subject: Re: Albus articles
[From Rick Marken (940707.0900)]
I would like to thank Tom Bourbon (940707.0813) for posting the summary ofthe Albus model. How in the world did you do it, Tom? I started gettingnauseous while studying the system diagram and nearly barfed on the keyboardafter reading about the Sensory Processing, World Model, Value Judgment andBehavior Generation components of the "model". I hope you were able to get holdof an ice-cold glass of 7-Up after posting this stuff. You're a better man thanI.
Now that I have recovered my sea legs I can see that the Albus model is anexcellent example of what PCT is NOT. Albus throws in every misconception inthe book, each one neatly summarized in the description of the function of eachbox in the model -- SP, WM, VJ, BG. It's tough to decide where to start dealingwith all these juicy misconceptions -- each one part of the bread and butter(I've become partial to scones and clotted cream after the England trip) ofcurrent behavioral science dogma. But I think your suggestion is probably thebest:
> I encourage the objector(s) to ... implement a single-loop version ofan Albus model, with all of the elements, connections, and processes that areshown in the figure I reproduced above.
Of course, maybe the Albus model can't do anything as dumb as keep a cursoraligned with a target; maybe the only thing the Albus model can do is act"intelligently". That would sure show us, wouldn't it? ;-) (NB. I AMbeing cute and judgmental).
> I am not being "cute" or "judgmental" in this offer;
Sure. Sure. I know you. You're the killjoy who would keep Uri Geller fromtouching the fork before doing his great psychokinetic fork bending routine.People love the Albus-type stuff because it uses all the right buzzwords. Nowyou wanna go and spoil it with all that science stuff. You're just no fun (forpsychological charlatans, anyway) but I can't wait to see you again in Durango;-).
Best Rick
Date: Thu Jul 07, 1994 3:18 pm PST
Subject: Re: Albus articles. System design is easy!
[From Bill Powers (940707.1100 MDT)] Tom Bourbon (940707.0813)
Thanks for that review of Albus: you have done us a service.
How easy modeling is when you don't have to make anything actually work!That "world model" is very handy; it contains all the properties of the worldas well as everything that can happen in it, and all that's needed to sense itis one little arrow that goes over to a box called sensory processing, with an"update" arrow going the other way. I wish somebody had told me that this isall there is to it. Out of this 40 year project I could have saved 39 years and11 months.
Actually, this is just the sort of system diagram that certain kinds ofexecutives like to draw up.
"What's your big problem?", they say, whipping out a sheet of newsprint anda dry-marker. "Watch, it's easy: just make a world model, connect it to asensory processor, hook that up to a value judgement box with all sorts of goodvalues and judgements in it, run the result over to the latest model ofBehavior Generator, stick in some more connections to the world model and someways of using it, and off you go! I don't see why _I_ have to do everythingaround here. I expect a progress report on my desk by Friday."
The poor engineer flunky, if he is wise, will say "Gee boss, that's a greatidea. You've really wrapped it all up here, all right. We'll get right on it."Then later, after he has worked out a simple control system and got it toreproduce some real behavior, he'll bring the results back and say "Gosh, youwere really right about that. Look at this schematic. Here's your sensoryprocessor, and we've got a signal coming in here that's exactly what youdescribed as a value, here here's the judgement that takes place just the wayyou said, and the output goes over here to your Behavior Generator that movesthe control stick. It all works beautifully, just the way you said it would.Only thing is, we need some advice on that world model part -- we're gettinggood behavior out of this, but you're going to have to give us some guidanceabout what the world model does. Maybe you could say that this system has an_implicit_ world model in it, so it really works just like your diagram.Amazing!"
You can always bullshit a bullshitter.
Best, Bill P.
Date: Thu Jul 07, 1994 3:18 pm PST
Subject: Re: PCT models and simulations
[Paul George 940707 10:30]
>[From Bill Powers (940706.1210 MDT)]
Heavily edited, with many good points omitted.
> Unless you are a very competent modeler, dabbling in thoughtexperiments can be the basis of delusions more easily than truth.
> Being, in fact, incompetent modelers, they {Psychologists} didn'trealize that an actual system organized in the way they proposed would haveproperties they hadn't anticipated, and would in fact blow itself up in a fewseconds, or else do nothing at all of any interest.
> If there's one thing any competent modeler knows, it's that anycomplex system design, unless it's thoroughly familiar and completely analyzed,is going to have important modes of operation that are different from what onethought they would be, and even more important modes of operation that aretotal surprises. Anybody can sit down and draw a fancy system diagram with allthe parts labelled with function-names, with nice-looking symmetries andimpressive patterns of repetition.
Amen. This is one reason why systems analysis is gravitating towards usingexecutable models and simulations for validating requirements and design. Inour world unforseen behavior can be deadly. Unfortunately, psychology has beenlimited by a tradition that experimentation on and vivisection of humans isunethical (Small minded of them ;-). There is no strong tradition ofexperimental proof of theory or models. Further, training in the soft sciencesoften contains little engineering, basic science, or mathematics. Statisticsare of course grossly misused and (I hope) misunderstood. Instead it hasessentially been a branch of philosophy (or theology ?? :-).
> One reason that PCT models tend to be simple is that we can't provethat any more complex models would actually behave as we imagine they would.All the models we talk about have been proven to work exactly as claimed, aswell as matching real behavior with acceptable accuracy. The idea of jumpinginto extremely complex behavioral problems while we still can't simulatesimpler behaviors is just ridiculous. We're building upward from a solidbase.
Forgive what may be misunderstanding, but from Dag's materials RickMarken's post yesterday, PCT derived from the concept that _People_ exhibitpurposeful behavior and perceptual control, not insects. That would seem (tooversimplify) to focus of the functions of the cortex, not the thalmus andnervous system. Thus, while your approach is good science and engineering, youhave validated only the PCT 'mechanism' (or one of them), not the PCT Theory.You (as a group) have demonstrated that simple positive feedback is sufficientfor low level neuro-muscular functions. You have not demonstrated that it isnecessary or sufficient for higher level behavior (to my extremely limitedknowledge). Thus Tom Bourbon's statement in an earlier post (as I recall) that"...we have found no need for more complex mechanisms" is not completelyaccurate.
In my mind this state of knowledge in no way decreases the value or utilityof the PCT theory, but fundamentally leaves it at the same level as otherbehavioral theories (including Albus) at the application level. It can bedemonstrated experimentally that operant conditioning works, but not _how_ itworks (i.e. the internal mechanism). Further we cannot generalize to say that_all_ behavior is just operant conditioning. The same kind of thing can be saidfor PCT behavioral control.
Paul George
Date: Fri Jul 08, 1994 9:32 am PST
Subject: Testing models
[From Bill Powers (940708.0905 MDT)] Paul George (940707.1030)
> Unfortunately, psychology has been limited by a tradition thatexperimentation on and vivisection of humans is unethical (Small minded of them;-). There is no strong tradition of experimental proof of theory ormodels.
What I was talking about was not matching models against behavior, butshowing that the model itself would behave as one claims it will behave. Thisis a much more serious problem. Look at Tom Bourbon's rendition of the Albusmodel. If you start at any arrowhead, for most of them you can trace half adozen or more ways of going around the system and ending up back where youstarted. Each distinct path is a feedback loop, involving differentcombinations of components. So you really have a large number of feedback loopssuperimposed and intertwined in this diagram, and if you ever constructed asimulation having this connectivity, all those loops would start interacting.It's extremely unlikely that such a simulation would act in the way Albusimagines that it would act.
What Albus has done is to go from his conception of how people might act toa diagram that describes all the pathways he noticed. But in drawing thediagram, he created, probably without realizing it, a large number of OTHERpathways. Some of these pathways would be negative feedback loops, but otherways of following effects around the system might pass through an even numberof sign-inversions and amount to positive feedback (unstable) loops. It thusbecomes impossible to get from the properties of the model back to the sameobservations that led Albus to draw the model in the first place. If a modelcould actually be constructed, one might be able to show the correspondencewith the real situation running in the direction from the observations to themodel -- but when the model ran, it would not reproduce the originalobservations. In fact, it would probably go into violent oscillations or lockup in states having nothing to do with real behavior, even the behaviors thatled Albus to draw this diagram.
This is the FIRST test that has to be applied to a model. Not a test to seeif the model's behavior would match real behavior, but a test to see if themodel would actually behave in the way you imagine it will behave, quiteindependently of the question of whether that behavior would match what a realsystem does.
All the models we use in experimental PCT actually behave in the ways weclaim they will behave, without unwanted or unexpected modes of behavior. Anyproposition that claims to be an actual model has to have this property. Veryfew models of behavior to be found in the literature do.
-------------------
> from Dag's materials Rick Marken's post yesterday, PCT derived fromthe concept that _People_ exhibit purposeful behavior and perceptual control,not insects.
The basic idea was that people exhibit _control_ behavior. From the modelof control behavior (out of engineering control theory, by Weiner) it could beseen that the role of the reference signal fits the properties of what weinformally call purposes, intentions, or goals.
> You (as a group) have demonstrated that simple positive feedback issufficient for low level neuro-muscular functions.
Typo, I hope: simple _negative_ feedback. Positive feedback is the oppositeof control.
> You have not demonstrated that it is necessary or sufficient forhigher level behavior (to my extremely limited knowledge).
We have a few examples, but none instrumented or quantitatively modeled.There are many informal examples of higher-level control, but the problem withmodeling it is that we don't know how to model the perceptual functionsinvolved. We'll get there eventually.
> It can be demonstrated experimentally that operant conditioning works,but not _how_ it works (i.e. the internal mechanism).
I have modeled operant conditioning using control systems, and the modelaccurately reproduced some data obtained by others. One day I'll resurrect thatmodel and post it. However, the model described by Skinner et. al. says onlythat reinforcement increases the probability of a behavior, increases theamount of behavior, or maintains a given behavior. When you try to turn thosewords into a working model, you end up with something that doesn't behave theway the real organisms do. The only model that does work treats the rate ofreinforcement as a controlled variable, with a reference signal for the desiredrate inside the organism, and the error adjusting the rate of behavior to bringthe actual rate of reinforcement toward the desired rate. And from this model,we conclude that "operant conditioning" is mostly ordinary control behavior,misinterpreted. So it isn't operant conditioning that works; it's control thatworks.
---------------------------
In the realms where we haven't experimentally modeled the PCT organization,we have to treat the PCT model as an explanatory construct. Since that is whatmost other models are (and nothing more), we can then ask how realistic thisexplanation is, in terms of what we observe and what we experience, incomparison with what other models tell us. So far, those who have learned allthe ins and outs of the PCT model say that this is the only model they haveever seen that represents their own behavior to themselves, and other people'sbehavior, in a believable form. Getting to that point of understanding takesrather a long time, a couple of years of asking questions and seeing howobjections are met. I hope you stay the course.
Best, Bill P.
Date: Fri Jul 08, 1994 12:06 pm PST
Subject: Re: Testing models
From Tom Bourbon [940708.1205] >[Bill Powers (940708.0905MDT)]
A taboo against vivisection of humans isn't the problem. In fact, where ithas been performed on other species (in deafferentation studies, for example),the results have not helped increase our understanding of behavior. To thecontrary, those studies have contributed to the great social danger thatfollows from most behavioral research and theory.
Bill P:
> What I was talking about was not matching models against behavior, butshowing that the model itself would behave as one claims it . . . . do withreal behavior, even the behaviors that led Albus to draw this diagram.
Yes, Bill. This was exactly what I saw when I began reading Albus again,and what I saw reminded me of why I had dismissed Albus back in 1979. It looksto me as though he noticed what, to him, seemed to be certain kinds of actionsand results that people produced, then he reified those observations into"explanatory constructs" (X is intelligent behavior; intelligence causesintelligent behavior; intelligence causes X), then he strung those constructstogether into a densely interconnected flow chart.
In 1979, while I was still a PCT neophyte, it was obvious that he hadcreated a typical behavioral-life science flow chart, not a model that wouldbehave, or if it would behave, there was no way he could possibly know inadvance what it would do. In the post where I reviewed his papers, that thoughtwas behind my remark that I could easily understand how people might carve outbits and pieces from the Albus densely-interconnected flow chart (the ADIFC)and use those pieces to build or program a working model, but their workingmodel would in no way be a working ADIFC.
> This is the FIRST test that has to be applied to a model. Not a. ..
Yes. Yes. Yes. And that is one of the most difficult ideas for us to getacross to our readers, editors, reviewers and critics. At this level, it is nota matter of PCT vs any other theory; it is a matter of whether _any_ particulartheory or model would really produce the results its proponents claim. This isan issue that cuts right to, and through, the heart of nearly every singlealleged theory and model in the behavioral-social-neuro- sciences. If thosepeople tested their "models" in the way Bill is describing here, practicallyall of them would come up lacking and would either be modified (into PCT-likemodels, by the way -- control models), or they would be tossed aside. But howlikely is it that practitioners of those "sciences" would willingly dismantleand close down their sciences? Don't count on it -- there are too many millionsof dollars in research funds waiting for them -- yet another of those gravesocial dangers to which I allude from time to time.
Later, Tom
Date: Fri Jul 08, 1994 1:54 pm PST
Subject: Re: Testing Models
[From Rick Marken (949708.1310)] Tom Bourbon (940708.1205) --
> At this level [of seeing whether the model works as expected], it isnot a matter of PCT vs any other theory; it is a matter of whether _any_particular theory or model would really produce the results its proponentsclaim. This is an issue that cuts right to, and through, the heart of nearlyevery single alleged theory and model in the behavioral-social-neuro- sciences.If those people tested their "models" in the way Bill is describing here,practically all of them would come up lacking and would either be modified(into PCT-like models, by the way -- control models), or they would be tossedaside.
It's just this kind of extremist talk that alienates sober,well-established, prominent behavioral-social-neuro-scientists from PCT. Why,you're saying that the entire edifice of the behavioral-social-neuro-sciencesis built on sand; that it's all a crock. Now who in the world is going tobelieve that (other than fools like you and me who have actually gone into thelab and found out for ourselves).
Excellent post, Tom!
Best Rick
Date: Fri Jul 08, 1994 3:44 pm PST
Subject: Re: PCT models
[Paul George 940709 11:30] >[Bill Leach 940707.23:04EST(EDT)]
> Paul; Your discussions have been of value to me personally. Thepoints you raise are resulting in some very illuminating responses.
Thanks. In truth I pointed out the Albus article in the spirit of throwingraw meat to wolves. The response of 'the masters':-) has been both informativeand amusing. Defending and attacking expose what the proponents of a theoryconsider most important with amazing rapidity.
On the whole I have found PCT itself to be very solid, and capable ofexplaining a great deal. At its philosophical level I find it appealing,certainly in comparison to behaviorist purists (Though I have often thoughtSkinner was grossly misunderstood) or some of the humanist school (dates mystudies, doesn't it).
But theory has two uses, understanding and application. As Jeff Vancouveris quoted as saying - Psychologists are trying to do something useful. Iinterpret this to mean apply theory to helping people, society, or personallife.
>From Tom Bourbon [940707.1729]
> ...one reason I say most claims to knowledge by psychology aredangerous is that innocent people might be tempted to believe those claims --some might even go so far as to appeal to psychological "knowledge" when theyattempt to establish new social policies, laws, and the like.
>From Tom Bourbon [940708.0816]
> All there is to support the majority of "theories" and "models" in thebehavioral-social-neuro- sciences is their apparent plausibility -- they soundas though they could be true.
It goes beyond plausibility to the realm of utility. Behaviorists may notbe able to provide a working model of the mechanism, but can give heuristicsfor how positive or negative reinforcement affect observed behavior andlearning. Gestalt psychology has observed and defined certain stages ofcognitive development, even if they don't provide the mechanical underpinnings.Priests and ministers can help people with life and make them more comfortable(and the reverse in spades!), even if metaphysics and theology are inherentlyun-provable. Alchemists were able to make chemical compounds even if theearth/air/fire/water element were not the correct model. A hot air balloon willwork regardless whether we use gas density or the Aristotelian concept of adesire towards heaven as an explaination. The point is all these theories areor were useful to some extent. They could also cause people to come to wildlywrong conclusions under certain circumstances. It is always a question of fact.The question is not "is it right" but rather "does it work".
>[From Rick Marken (940707.1415)] "Out on a limb"
>From Tom Bourbon [940707.1729]
> Rick Once you demonstrate that a higher level behavior is a controlledvariable, then you know you need a control model to explain it. A control modelis necessary because you are modelling the phenomenon of control.... -- butwhen behavior involves control, you can only model it with a controlmodel.
This I feel is an oversimplification. A control model is required tocapture the _control component_ of the behavior. Further PCT has not (to myknowledge) that higher level behavior _is_ a "a controlled variable". Ipersonally suspect it involves control of sets of controlled perceptualvariables. It also IMHO involves sets of controlled reference variables,reference values, and some mechanism to determine from memory and/orextrapolation what variables of both types are or are going to beimportant.
{out of order}
> The answer, according to PCT, is that higher level behavior (like allbehavior) is _a controlled result of action_.
I am not sure that control phenomena is all that exists. There arecomponents that appear reactive. (I didn't fully follow the thread on alerting,other than to observe the apparent response that the behavior could be producedby a sufficiently fast control loop). When a hand is burned, the arm will jerkback. It is my understanding that this does not involve the brain, just neuralreflex at the spine. I seem to recall something on the order that the opticnerve of a frog transmits no signal to the visual cortex until a moving objectlike a fly enters the field of vision.
Having a continuously active control loop monitoring every possible controlvariable seems very inefficient. Some inputs are discontinuous or periodic.Bill Powers statements recently notwithstanding, a control loop can stop, or atleast idle. Computer systems often use something called and interrupt chain. Ifno sensory inputs (events) are currently active (one of the results of an errorsignal existing can be the generation of a discrete input to another loop) thecontrol loop hibernates. When woken up, one kind of control loop determineswhat kind of event (or control variable) was populated or went out or range andactivates the appropriate control loop to get things back in order.
> So the idea that PCT can handle simple behavior while the jury isstill out on its ability to handle higher level behaviors really misses thepoint completely. Simple vs higher level is not the dimension on which PCT isdiscriminated from other theories of behavior; the dimension is "generatedoutputs vs control".
What I have noted in most discussions, particularly that about Albus, isthe focus upon how "PCT is discriminated from other theories of behavior". Youfocus upon the differences, not the commonalities or how the approaches mightbe blended into a more powerful whole. It's kind of like "well, that's not whatwe're talking about, or it isn't the same point of view, or it doesn't addresssomething we do. Now having disposed of that, we can go back to our own worldand what we're doing."
The major difference I see between PCT proponents and others is thattraditional psychology and sociology looks at behavior from the outside of theorganism, while PCT focuses on the inside. Psychology has said 'We observe thefollowing[...].. What can we say about predicting it or understanding therelationships between patterns of apparent events and observable actions of theorganism?' PCT says 'can we identify something being controlled internally anda negative feedback{ack Rick} relationship? Can we simulate it?' (yes it's moreinvolved than that). It is kind of like the difference between a 'top down' and'bottom up' approach to analysis. Both have their pitfalls.
I would challenge you to try to design or understand a distributed realtime system at the assembly language or register level. It _can_ be done, butnot by many. But higher level constructs such as 'tasks', 'messages', datastructures are useful for understanding and constructing the application, evenif it does ultimately reduce to bit twiddling. Similarly, to understand thoughtusing only the mechanics of dendrite firing is at best difficult.
To say "if they can't do the simple stuff why should we pay attention toanything you have to say at higher levels" is at best unfair and at worstdisingenuous. Their models do not purport to address that level of detail.Frame of reference and scale of reference are important concepts to keep inmind.
Ultimately I suspect PCT will meet other approaches somewhere in themiddle, providing both mechanism and greater understanding of the parameters ofsome kinds of human behavior. When you can build an executable control modelshowing how one puts on a sweater and/or adjusts a thermostat, and decidesbetween the options, I will be convinced. A skin temperature control variableis not quite sufficient. Ah well, +sufficient unto the day be the troublesthereof'.
Date: Fri Jul 08, 1994 3:45 pm PST
Subject: Re: Testing models
[From Paul George 940708 15:00 EDT]
Can buy most of the post, but a couple nits:
[Bill Powers (940708.0905 MDT)]
> What Albus has done is to go from his conception of how people mightact to a diagram that describes all the pathways he noticed. But in drawing thediagram, he created, probably without realizing it, a large number of OTHERpathways. Some of these pathways would be negative feedback loops, but otherways of following effects around the system might pass through an even numberof sign-inversions and amount to positive feedback (unstable) loops.
This is true only if you model the diagram naively and permit all paths tobe traversed. If you put a state machine(s) or PCT control node(s) within thenodes, and/or any discrimination node which can rout different signals tovarious control or sensory nodes, then all paths are not possible. Behaviorcould be rather complex, but so is organismic behavior. Unstable systems (likeweather) may function within bounds (strange attractors). Besides, behavior(and life) is only a stable system within certain bounds.
A PCT control node must ignore all inputs that do not affect the controlledvariable, or more accurately never receives them. How is it that a HPCT systemassembles inputs, controlled variables, reference variables, and outputs, notto mention appropriate interconnections. All paths between nodes are notpermissible in the HPCT model either (of course you define internal connectionrules).
> And from this model, we conclude that "operant conditioning" is mostlyordinary control behavior, misinterpreted. So it isn't operant conditioningthat works; it's control that works.
Close. Control is _a_ mechanism which _can_ implement operant conditioning,given certain assumptions. How is the control variable and negative feedbacksignal set up? (I have a little trouble with 'Unseen Hand' arguments)
The organism says 'gee, I liked that! Now what can I do to get more of it?The reinforcement must exist before its rate may be controlled. The organismmust deduct a relationship between a behavior (or output of some sort) with thesensory input which is +reinforcement.+ Then a control loop must be set upwhich detects the rate (or more likely the presence). The control loop mustessentially generate a constant maxed error signal that intermittently goes tozero when reinforcement arrives or it's rate increases. A rate based controlvariable would essentially be a trend computation based upon an intermittentbinary signal (with a long period relative to the 'actions').
Several problems however: First, reinforcement is not a constant signal,andnot even the same signal. Second, the desired behavior is not always rewardedby the trainer. While an increase in the rate may be desired by the organism,the rate is in fact random or constant, and ultimately declining. Third, thebehavior persists after the reinforcement stops. The "...bring the actual rateof reinforcement toward the desired rate" doesn't occur. The problem is worsewith negative reinforcement. There the control issue becomes 'what behaviormust stop to reduce the unpleasant perception?". The concept of an error signalproducing action becomes a little convoluted in the OC case, or that ofMaslow's hierarchy. A 'happiness' or 'comfort' controlled variable is not muchmore convincing than a 'world model'..
Note: I don't really intend to have a lengthy discussion of the propertranslation of OC to a control model. However, demonstrating learning behaviorand OC using a 'dog training' type scenario would be an interesting test of theclaim that 'all behavior is control'. You would have to come up with amechanism for dynamic reconfiguration of control nodes and connections foradaptive purposes, which is a hard current computer science problem.
Not to worry, just because we don't know _how_ something works or we don'thave the technology to reproduce it doesn't mean it can't work. This is a basicproblem with simulation or modeling - the devil is in the details. We can't yetdesign and construct biological entities. PCTrs could be dismissed byneuro-physiologists because you can't describe how to build a PCT control nodeout of neurons and/or brain cells (I don't think).
>From Tom Bourbon [940708.1205]
> But how likely is it that practitioners of those "sciences" wouldwillingly dismantle and close down their sciences? Don't count on it -- thereare too many millions of dollars in research funds waiting for them -- yetanother of those grave social dangers to which I allude from time totime.
.a tissue paper dog chasing an asbestus cat through hell. Of course thesame can be said for any closely held belief, particularly one upon which one'sreputation or world view is based.
Looking forward to next week. - Paul George
Date: Fri Jul 08, 1994 4:32 pm PST
Subject: Re: PCT models
[From Tom Bourbon 940708.1653]
Should I wait until Monday, so I can cool down a bit? Nah!
>[Paul George 940709 11:30]
>>[Bill Leach 940707.23:04 EST(EDT)]
>> Paul; Your discussions have been of value to me personally. Thepoints you raise are resulting in some very illuminating responses.
> Thanks. In truth I pointed out the Albus article in the spirit ofthrowing raw meat to wolves. The response of 'the masters':-) has been bothinformative and amusing. Defending and attacking expose what the proponents ofa theory consider most important with amazing rapidity.
Glad to know we could keep you laughing for a while. For our part, we don'tusually suggest that people read and comment on published material unless wereally mean it. I don't especially like to play games like that. Call it asilly bias of mine about how I approach people and deal with them on the net,or anywhere else, for that matter.
Would you let us in on what you found the most amusing?
And would you maybe let us in on your personal assessment of Albus, nowthat we know you weren't playing straight with us for the past week and a half?Whatever your assessment may be, I guess we don't need to wait to see anyimplementations of a genuine Albus model by you, do we?
> On the whole I have found PCT itself to be very solid, and capable ofexplaining a great deal. At its philosophical level I find it appealing,certainly in comparison to behaviorist purists (Though I have often thoughtSkinner was grossly misunderstood) or some of the humanist school (dates mystudies, doesn't it).
I have a long-standing interest in Skinner and in the often-repeated claimthat he was grossly misunderstood. Could you give a few specific examples ofwhat you mean? (Or is there a hidden truth here, also? Are you just giving ussome more raw meat, looking for a little more amusement?)
Paul, what you are seeing in my reply (as I would interpret it using PCT)is a little hint of how a hierarchical control system feels, and what it does,when it discovers that someone has deliberately jacked around with perceptionsat the levels of system concepts and principles. You did that and I'm prettyhot.
> But theory has two uses, understanding and application. As JeffVancouver is quoted as saying - Psychologists are trying to do somethinguseful. I interpret this to mean apply theory to helping people, society, orpersonal life.
That's what I'm talking about also, when I say the behavioral sciences (andtheir near relatives in the sciences) are dangerous. As in:
>> From Tom Bourbon [940707.1729]
>> ...one reason I say most claims to knowledge by psychology aredangerous is that innocent people might be tempted to believe those claims --some might even go so far as to appeal to psychological "knowledge" when theyattempt to establish new social policies, laws, and the like.
>>From Tom Bourbon [940708.0816]
>> All there is to support the majority of "theories" and "models" inthe behavioral-social-neuro- sciences is their apparent plausibility -- theysound as though they could be true.
> It goes beyond plausibility to the realm of utility.
I'm open to specific examples.
> Behaviorists may not be able to provide a working model of themechanism, but can give heuristics for how positive or negative reinforcementaffect observed behavior and learning.
Fine. A technology and not a science. But that isn't where they stop withtheir own descriptions of what they do. Who was it that said behaviorism, morespecifically radical behaviorism, _is_ _the_ science of behavior?
> Gestalt psychology has observed and defined certain stages ofcognitive development, even if they don't provide the mechanicalunderpinnings.
Good enough, if we grant the veracity of those observations, which not alldevelopmentalists will do. But I thought science was supposed to entail morethan a few empirical observations.
> Priests and ministers can help people with life and make them morecomfortable (and the reverse in spades!), even if metaphysics and theology areinherently un-provable.
Yes, but do they then pretend to the title of "behavioralscientist?"
> Alchemists were able to make chemical compounds even if theearth/air/fire/water element were not the correct model. A hot air balloon willwork regardless whether we use gas density or the Aristotelian concept of adesire towards heaven as an explaination. The point is all these theories areor were useful to some extent. They could also cause people to come to wildlywrong conclusions under certain circumstances. It is always a question of fact.The question is not "is it right" but rather "does it work".
All well and good, but all irrelevant to the discussion about a science ofbehavior. Oh, but I forgot: you weren't really serious about that discussion,were you?
>>[From Rick Marken (940707.1415)] "Out on a limb"
>> From Tom Bourbon [940707.1729]
>>Rick
>> Once you demonstrate that a higher level behavior is a controlledvariable, then you know you need a control model to explain it. A control modelis necessary because you are modelling the phenomenon of control.... -- butwhen behavior involves control, you can only model it with a controlmodel.
> This I feel is an oversimplification. A control model is required tocapture the _control component_ of the behavior.
And you can demonstrate that a control model has no applicability beyondthat? Even Albus knew better than that. ("There is no such thing as _mere_motor behavior. . .. The intellect is _not_ something distinct frombehavior.")
> Further PCT has not (to my knowledge) that higher level behavior _is_a "a controlled variable".
Something is missing from your post. Did you want to say that PCT has not_said_ that higher level behavior is a controlled variable? Or that PCT has notdemonstrated ...? Or was it something else?
> I personally suspect it involves control of sets of controlledperceptual variables.
So do we, if by "it" you mean that what most people call higher levelbehavior, PCTers call behavior that controls higher level perceptions.
> It also IMHO involves sets of controlled reference variables,reference values, and some mechanism to determine from memory and/orextrapolation what variables of both types are or are going to beimportant.
Hmm. So you weren't _completely_ disingenuous when you appealed to Albus,after all. Time will tell, on this point. I would still like to know how youmight program and run the model for this kind of performance. Albus did notprovide that level of detail in his "outline." And I would like to know if youthink such a model would be a representation of how living systemsfunction.
> {out of order}
>> The answer, according to PCT, is that higher level behavior (likeall behavior) is _ a controlled result of action_.
> I am not sure that control phenomena is all that exists.
Certainly not. There are many natural phenomena other than control. Forexample, there are Cause--->Effect phenomena. There just aren't many (any?)examples of those in the behavior of living things.
> There are components that appear reactive. (I didn't fully follow thethread on alerting, other than to observe the apparent response that thebehavior could be produced by a sufficiently fast control loop). When a hand isburned, the arm will jerk back. It is my understanding that this does notinvolve the brain, just neural reflex at the spine. I seem to recall somethingon the order that the optic nerve of a frog transmits no signal to the visualcortex until a moving object like a fly enters the field of vision.
I was about to launch into a detailed reply to this point, but I forgotthat you still haven't read any of the PCT literature. In that literature thereare numerous discussions of examples in which a PCT analysis reveals that"components that appear reactive" rarely are purely reactive.
> Having a continuously active control loop monitoring every possiblecontrol variable seems very inefficient.
Not every variable -- every controlled perception.
> Some inputs are discontinuous or periodic. Bill Powers statementsrecently notwithstanding, a control loop can stop, or at least idle.
Or do they _seem_ to do that?
> Computer systems often use something called and interrupt chain. If nosensory inputs (events) are currently active (one of the results of an errorsignal existing can be the generation of a discrete input to another loop) thecontrol loop hibernates.
Sure a computer system can use that procedure; but we are trying to studyand model living things and they are a whole different breed of cat.
>> So the idea that PCT can handle simple behavior while the jury isstill out on its ability to handle higher level behaviors really misses thepoint completely. Simple vs higher level is not the dimension on which PCT isdiscriminated from other theories of behavior; the dimension is "generatedoutputs vs control".
> What I have noted in most discussions, particularly that about Albus,is the focus upon how "PCT is discriminated from other theories of behavior".You focus upon the differences, not the commonalities or how the approachesmight be blended into a more powerful whole.
Please. At the start of this post, you said you weren't really seriouswhen you appeared on this net with the following post:
======================
P. George, 23 June 1994:
I have been lurking for a while listening to the PCT debates, and think youmay be using too simplistic a view of a control system, which is biasing yourdiscussion (in fairness I haven't had access to the books/papers on thesubject, just posts on BPR_L by Dag).
Sophisticated control systems don't use reference variables (e.g. setpointsor alarm thresholds) they use reference _models_ (reflected in control logic).In a sense we have a continuously running simulation of the 'real' world towhich we compare our perceptions (sensor inputs). As PCT stresses we alsofilter our perceptions based upon our current model, trying to separate signalfrom noise.
The basic components of this view of a control system are SensoryProcessing (filter/transducer), Value Judgement (comparator), WorldModel/database, and Behavior Generator. Part of the model is some set of goalsor desirable states we wish to maintain or approach. We modify the model tobetter reflect 'outside reality' (as perceived) so that we can better predictthe results of our actions (actuator outputs).
For a good exposition of this model I suggest:
"Outline for a Theory of Intelligence" James S. Albus IEEE Transactions onSystems, Man, and Cybernetics, vol 21 #3, May/June 1991, p 473-509. IEEE log #9042583 about the Albus article.
========================
It looks to me as though _you_ directed the focus to ""how "PCT isdiscriminated from other theories of behavior"."" Please don't tell us that_we_ pressed that focus upon you. You appeared making strong assertions ofdifferences, which you now say were not really serious but were in fun, and nowyou say we have focused on the differences. Come on, Paul.
> It's kind of like "well, that's not what we're talking about, or itisn't the same point of view, or it doesn't address something we do. Now havingdisposed of that, we can go back to our own world and what we're doing."
Another of your little jokes, looking for some more laughs? I'm more thanmildly amused, myself.
. . .
> I would challenge you to try to design or understand a distributedreal time system at the assembly language or register level. It _can_ be done,but not by many.
Fine. And I repeat my invitation (not a challenge) for you to take asingle-loop, fully-interconnected, Albus model and show that it duplicates theperformance of a person and a PCT model on a simple control task. Let's startthere, Paul.
> But higher level constructs such as 'tasks', 'messages', datastructures are useful for understanding and constructing the application, evenif it does ultimately reduce to bit twiddling. Similarly, to understand thoughtusing only the mechanics of dendrite firing is at best difficult.
The simple stuff first, please. My invitation still stands.
> To say "if they can't do the simple stuff why should we pay attentionto anything you have to say at higher levels" is at best unfair and at worstdisingenuous.
Did someone here say that? I believe I said something to the effect that ifthe causal mechanisms assumed in a theory or model of behavior can't produceaccurate and reliable results in simulations of the "mere" or "simple" kinds ofbehavior, then there is no possibility that that causal mechanism can explainand predict "higher" or "more complex" behavior. Maybe I'm wrong. Can you tellme where my reasoning has gone astray?
> Their models do not purport to address that level of detail. Frame ofreference and scale of reference are important concepts to keep in mind.
Don't they, now? I seem to see a literature _filled_ with assertions aboutthe details of how behavior is produced -- it is planned, commanded,controlled, programmed, and so on. Those untested and unworkable assertionscome from people who purport to have a theory of complex behavior -- theimportant word is "behavior." I don't put that word into their writings, theydo. If they speak of behavior, and further if they say they can explain complexbehavior, then I expect to see at least _a little_ evidence that their assumedcausal mechanisms can work. I guess there is very little I can do about thefact that this attitude of mine strikes you as disingenuous and amusing.Them's the breaks.
> Ultimately I suspect PCT will meet other approaches somewhere in themiddle, providing both mechanism and greater understanding of the parameters ofsome kinds of human behavior.
Go ahead and suspect all you want. I _had_ hoped that you might actuallyjoin in and help with the project, but now I know you were only playing.
> When you can build an executable control model showing how one puts ona sweater and/or adjusts a thermostat, and decides between the options, I willbe convinced. A skin temperature control variable is not quite sufficient. Ahwell, +sufficient unto the day be the troubles thereof'.
Ah, well, since all we have ever written about and modeled is skintemperature, I guess we should simply leave the field to the real scientists --the ones looking at the Really Big Questions. (Read some of the PCT work beforeyou say such things about it, Paul. Don't keep speaking out ofignorance.)
I am happy that you suggested we, "build an executable control modelshowing how one . . . adjusts a thermostat." That happens to be one of themodels I will demonstrate at the upcoming meeting of the CSG, in Colorado. It'snice to know there is at least one person in the world who might be interestedin the results.
Oh, and by the way. Would you tell us about some of the people who havealready built the models for putting on a sweater and adjusting a thermostat?Or has no one done that and you are just urging us to be the first? If theyhaven't already done those things, what _have_ they done that has convinced youthey are on the right track?
Should I send this flaming piece? Probably not, but what the heck.
Later, Tom
Date: Fri Jul 08, 1994 5:43 pm PST
Subject: Look before you leap.
[From Bill Powers (940708.1830 MDT)] Paul George (940709.1130)
> Thanks. In truth I pointed out the Albus article in the spirit ofthrowing raw meat to wolves. The response of 'the masters':-) has been bothinformative and amusing. Defending and attacking expose what the proponents ofa theory consider most important with amazing rapidity.
Sorry -- they expose what you THINK they consider most important, theexposure occurring more rapidly the more willing one is to jump to simplisticconclusions.
Paul, I really do urge you to get up to speed with the PCT literature. Ifyou keep making comments of the kind you've been offering before pausing tofind out what you're talking about, you will have generated such a backlog ofembarrassing misinterpretations that you would have to be Jesus Christ to admitto them. Why not bite the bullet and learn what PCT has to say? Come on, itwon't kill you.
Best, Bill P.
Date: Fri Jul 08, 1994 5:52 pm PST
Subject: Re: PCT models
[From Dag Forssell (940708 1800)]
On June 16, I posted on BPR-L: (Business Process Reengineering -L)
------------------------------
> Bill Powers writes regularly on the Control System GroupList:
CSG-L
> To subscribe and learn of your options, send
> Subscribe CSG-L Firstname Lastname
> help
> info refcard
> query csg-l
> to: LISTSERV@VMD.CSO.UIUC.EDU
> CSG-L volume is about 300 KB per week, against BPR-L 300 KB per month.Many posts are long and technical, but sincere questions, no matter how simpleare welcome and get careful answers. Listen in, please read the literaturefirst and introduce yourself when ready. See you there.
-----------------------------
Paul, you lurked for an entire week, before you emerged from the shadowswithout having read the literature. You still have not.
Now you say:
> In truth I pointed out the Albus article in the spirit of throwing rawmeat to wolves. The response of 'the masters':-) has been both informative andamusing. Defending and attacking expose what the proponents of a theoryconsider most important with amazing rapidity.
You have asked insincere questions and received careful answers.
Now you say to Rick:
> This I feel is an oversimplification. A control model is required tocapture the _control component_ of the behavior. Further PCT has not (to myknowledge) [shown??] that higher level behavior _is_ a "a controlled variable".I personally suspect it involves control of sets of controlled perceptualvariables. It also IMHO involves sets of controlled reference variables,reference values, and some mechanism to determine from memory and/orextrapolation what variables of both types are or are going to beimportant.
and
> I am not sure that control phenomena is all that exists. There arecomponents that appear reactive. (I didn't fully follow the thread on alerting,other than to observe the apparent response that the behavior could be producedby a sufficiently fast control loop). When a hand is burned, the arm will jerkback. It is my understanding that this does not involve the brain, just neuralreflex at the spine. I seem to recall something on the order that the opticnerve of a frog transmits no signal to the visual cortex until a moving objectlike a fly enters the field of vision.
You are arguing from a standpoint of rather complete ignorance of PCT. PCTexplains the reflexes you mention. You have been shown great courtesy in theform of sincere answers to your "throwing raw meat to wolves". I would like toperceive you returning the favor, by asking sincere questions. Sincere based ona real desire to understand.
> Looking forward to next week.
Please help make this mutual. Read the literature so you can learn from theexchanges, not just entertain yourself by jerking sincere PCT advocatesaround.
Best, Dag
Date: Fri Jul 08, 1994 9:35 pm PST
Subject: Re: PCT models
<[Bill Leach 940708.21:38 EST(EDT)] >[Paul George 94070911:30]
Unfortunately both Ed Ford and Dag Forssell appear to have been very busyof late. Both of these people are "practicing" PCT in counseling or teaching.The both spend a great deal of time interacting with people that 1) have neverheard of PCT and 2) could care less (at least initially).
I believe that Ed might tell you something like: "The very best counselorsand therapists actually employ many principles of PCT in their work but would(and usually do) fervently deny doing so."
> It goes beyond plausibility to the realm of utility. Behaviorists maynot be able to provide a working model of the mechanism, but can giveheuristics for how positive or negative reinforcement affect observed behaviorand learning.
The real root of the problem here is precisely what Tom is talkingabout:
Yes, INDEEDY, they provide heuristics and such for various behavior. Theyhave names for about every "abnormal" behavior condition that exists and inmany cases they have "treatments" designed to "correct" the behavior. Thereason that this is actually dangerous to the world is that, if the basic PCTconcept is right -- and there is compelling evidence that the basic concept isindeed right --, then these behaviouralists are treating symptoms and notcauses. This can work (and indeed sometimes actually might work) but theproblem in the realm of psychology (as predicted by PCT) is that any given setof symptoms could be caused by an almost infinite range of "causes".
This could be compared to how a medical doctor operates. The medical fieldis still almost a soft science itself but at least doctors actually test forreal, verifiable physical conditions prior to responding to some symptom set.As I am sure that you are aware, failure to do so would lose a large percentageof patients (again, same symptoms can be produced by wildly varying causes,each of which requires a different treatment methodology).
Basically, this is how PCT and the other behavioral "sciences" are related.PCT recognizes "behavior" as a symptom. PCT recognizes that a statement such as83.5% of the test subjects reacted "such and such way" to a stimulus, is onlysaying that 83.5% of the test subject were controlling the same reference thesame way (and that actually should be worded ... roughly the same reference inroughly the same way -- since in most studies many conditions important to thestudy are neither controlled nor monitored). The question that PCT both raisesand is concerned with is... what about the other 16.5%? Are they not realpeople too? An 80 or 90% correlation sounds good (and it is rarely actuallythat high in any psych studies even if the study says otherwise) but it ispretty lousey.
How would one feel if the laws governing hydraulics and mechanics (such aspertain to the coefficient of friction) only applied about 90% of the time.Thus, roughly every tenth time that you applied the brakes on your automobilethey just did not work -- of course the same would be true for everyone else'sbrakes too (and we think that we have carnage on our hwys now!).
> Priests and ministers can help people with life and make them morecomfortable (and the reverse in spades!), even if metaphysics ...
And the logical extensions to the theory of PCT can help to explain bothresults and why.
> Alchemists were able to make chemical compounds even if theearth/air/fire/water element were not the correct model.
Yes and it is quite likely that Alchemists killed a fair number of peopletoo (including themselves). There is nothing wrong with doing something usefulusing a script even if the theory is incorrect. Where the error occurs is whenone tries to generalize and apply the incorrect theory where the theory'serror(s) are significant.
All of the sciences have had their "bad days" and rejected the correctapproach in a big way (most do something similar in a small way all the time ofcourse). The issue with the behavioral sciences right now is that they are only"making noise" concerning application of scientific method to their work.Behavioral scientists are not rigorous. They basically all run around sayingthat human behavior is too complex to be able to prove a theory. Havingaccepted that idea, they then refuse to either examine hard science data orhave such principle based tests applied to their own theories.
This might not be so bad except that like the doctor, they are literallykilling their patients and are providing "guidance" to policy makers, lawmakers, leaders and managers that is just plain wrong. Fortunately (or maybethat should be unfortunately) people seldom actually follow the advice fully(including the promoters themselves).
The behavioral sciences is a field full of inconsistent rules andexceptions or theories so vague that one can not actually concludeanything.
You can probably no more meld PCT to other theories than you can design oneof your plant control systems without using negative feedback control (thinkingabout that, you actually can control some industrial processes without explicitnegative feedback control though it tends to make a failure pronesystem).
> This I feel is an oversimplification. A control model is required tocapture the _control component_ of the behavior. Further PCT has not (to myknowledge) that higher level behavior _is_ a "a controlled variable".
And this is the essence of misunderstanding of PCT (I believe). What PCT asa theory states is that BEHAVIOR IS THE CONTROL OF PERCEPTION. That's it,that's all of it and there is pretty sound reasoning (in addition to whatactual testing has been performed) to be able to conclude that there need notnecessarily be anything else.
> Having a continuously active control loop monitoring every possiblecontrol variable seems very inefficient. Some inputs are discontinuous orperiodic.
I don't know that the theory says that this is the case for all situationsbut I doubt that the opinion that it "seems very inefficient" is a particularlycompelling reason to suspect that it is not the situation.
I suspect that you are possibly confusing "conscious" control with controlin general. In some of your plant systems, the control systems do not stopcontrolling when they do not have anything to do but rather they control for noaction. In your distributed processing systems, it is quite common forcontrollers to operate for extended periods without reporting or requestingfrom their host and indeed in some such systems no report (other than maybe aperiodic sanity check) is made unless there is either a control failure or"perception" that there may soon be one. This is still control and it is alsocontrol in the PCT sense too.
> What I have noted in most discussions, particularly that about Albus,is the focus upon how "PCT is discriminated from other theories of behavior".You focus upon the differences, not the commonalities or how the approachesmight be blended into a more powerful whole.
But the difference IS significant. If behavior is "response to stimulus"(as it can often appear to be when the stimulus disturbs a controlledperception) as opposed to control of perception then PCT is utterly wrong andhas nothing to offer the world about the operation of living organisms. I amsure that a great deal has been learned by the PCT folks in studying otherbehavioral scientists' work, indeed they often provide the "proof" of thefailure of their theories within their own reports.
> The major difference I see between PCT proponents and others is thattraditional psychology and sociology looks at behavior from the outside of theorganism, while PCT focuses on the inside. Psychology has said 'We observe thefollowing[...].. What can we say about predicting it or understanding therelationships between patterns of apparent events and observable actions of theorganism?' PCT says 'can we identify something being controlled internally anda negative feedback{ack Rick} relationship? Can we simulate it?' (yes it's moreinvolved than that).
Unfortunately, the problem with the observation is that if such reasoningis based upon a faulty model of the fundamental operation of living beings theconclusions will at best be true for only a very limited (and probablyunspecified) set of conditions -- which is exactly the case now.
> It is kind of like the difference between a 'top down' and 'bottom up'approach to analysis. Both have their pitfalls.
This sounds like a valid argument but it is not (in my opinion anyway)."Top down" and "bottom up" both deal with exactly the same knowledge andexactly the same understanding of operational details.
"Top down" and "bottom up" are not equivalent to "Alchemy" and "Chemistry"nor is other "behavioral science" related to "PCT" that same way. If the PCTfield were large enough then indeed there might be people working bothapproaches within PCT but I don't think that is the case right now.
> I would challenge you to try to design or understand a distributedreal time system at the assembly language or register level. It _can_ be done,but not by many.
Again, the comparison is not in my mind fully valid. Someone somewhere didactually fully understand the details of hardware operation of such systems.The libraries (or OS access modules) for example are written by people thathave intimate knowledge of the hardware. They precisely specify the callingsequence for their use and "hide" the gory details but the details areknown.
-bill
Date: Sat Jul 09, 1994 12:12 am PST
Subject: Re: Testing models
<[Bill Leach 940708.23:01)] >[Paul George 940708 15:00]
I probably should do a little more controlling for accomplishing some workaround here but seems I can't resist. <G>
> This is true only if you model the diagram naively and permit allpaths to be traversed. If you put a state machine(s) or PCT control node(s)within the nodes, and/or any discrimination node which can rout differentsignals to various control or sensory nodes, then all paths are notpossible.
This supports what has been said. Something would have to CONTROL the"model" so why bother with that model?
> A PCT control node must ignore all inputs that do not affect thecontrolled variable, or more accurately never receives them.
Interesting assertion. Why does this have to be true?
Pavlov's dog is explainable through PCT and pleasantly enough, theexplaination is even capable of explaining why the dog does not always seek thereward.
> The organism says 'gee, I liked ...
You really do need to read some of the publications in PCT. My impressionis that you have tried to define how the control system must operation and thenchallenge your own design. I am not qualified to really discuss in depth howthis "reinforcement" process actually works but am willing to note that withall of the claims of its proponents there are (or were) eminent psychiatrists(Frankl comes to mind) that refuted the validity of the theories (though theydid not have any replacement theories).
-bill
Date: Sat Jul 09, 1994 12:30 am PST
Subject: Re: PCT models
<[Bill Leach 940708.21:11 EST(EDT)] >Tom Bourbon[940708.0816]
I probably should have phrased that in more precise terms. PCT is notproven true in the sense that it can be taken as a physical "law" (such as likecharges repel -- and even that is not a good example), ... yet.
The choice of the word "plausible" was poor also in that people (at leastin the "soft" sciences and the general public) tend to view plausible asmeaning "reasonable" rather than the idea that there is a actual body ofphysical evidence related through consistent laws or theories that can beexamined objectively.
Basically, what I hear again and again on the CSG-L is that HPCT and theconjectures concerning it are hypothesis not theory. However, that hypothesishas the advantage over other behavioral "theories" in that it has not failedany serious challenge. The problem being that we don't KNOW that HPCT iscapable of generating the observed human behavior but rather it is plausiblethat it could do so.
-bill
Date: Sat Jul 09, 1994 6:44 am PST
Subject: No control without feedback
[From Bill Powers (940709.0740 MDT)]
Bill Leach (940708.xxx) --
Thanks for the morale-building private post, Bill.
In trying to get through to Paul George you make a whole series of lovelyobservations. In some places you could make an even stronger case:
> You can probably no more meld PCT to other theories than you candesign one of your plant control systems without using negative feedbackcontrol (thinking about that, you actually can control some industrialprocesses without explicit negative feedback control though it tends to make afailure prone system).
I think we can say flatly that there is no way to design a control system(one that creates consistent results by acting on and in a real environment)without feedback. To prove this, all you have to do is keep a sharp lookout forthe feedback that is ALWAYS present when continuing regularities areseen.
To control anything literally without feedback, you would have to set upthe situation so the person or machine doing the controlling could issuecommands that affected the processes in the plant, BUT WAS NEVER ALLOWED TOOBSERVE THE RESULTS either while issuing the commands or at any later time.Just try to imagine how we could control ANYTHING under those conditions. Youwould remain blissfully unaware of failures, wouldn't you?
Or go even further: give a person or machine the ability to perform actsthat affect the plant, and the ability to sense the resulting operation of theplant, but eliminate any conception of the desired operation of the plant:whatever it is seen to do is OK, no error implied, no preferred kind of resultimplied. The perceptual information is just for the entertainment of theobserver, and does not call for any change in the action. Would anyone callthis control?
Sometimes engineers think they are really building an open-loop system.They calculate every contingency, build everything to micron specifications,double and triple check every component and every physical theory involved, andeliminate every possible source of disturbance. Then they send the design tothe production department. But do you think for an instant that such anengineer (if sane) would walk away without seeing what this system actuallydoes? No, what you find is that the engineer is standing by anxiously with atwiddle-stick, and spends a lot of time trimming and adjusting and bendingthings a little until the system is actually producing the effects that theengineer -- the control system -- wanted to see. And even after that, no saneengineer would seal the system to protect it from further alterations and signthe guarantee that it will work forever. Far more likely, the engineer willleave behind a thick book of maintenance and calibration procedures, which heexpects will be followed daily, preferably hourly, and certainly at leastmonthly, by the person using the system. The engineer knows that in theenvironment of the real Turing machine, there is a kid with a spraycan who liketo spray dots onto any tapes he comes across.
There is ALWAYS a control system involved, even if it's only the person whois using the open-loop device. Without that control system, only an imitationof control is possible. If the control system does not act, sooner or later theopen-loop system will wander blindly off into stupid actions that not only failto accomplish the original intention, but actively produce results other thanthose intended.
The concept of an open-loop automaton is a vestige of 18th- and19th-century engineering, from the time when people felt that the universe wasa divine clockwork mechanism that, once God had set it in motion, would runforever with divine precision. In fact, natural processes wander randomly,never exactly repeating and always doing some sort of random walk toward nogoal-state at all. Chaos prevails. Everything disturbs everything else. Themore closely you look at natural processes, the more obviously the sense oforderliness is violated. The basins of attraction in nature are shallow andsubject to disturbance.
In this universe only a feedback control system can create the same resultsover and over indefinitely, defying both chaos and probability. Only actionsadjusted by purposes can create permanent regularity while the purposelasts.
Best, Bill P.
Date: Sat Jul 09, 1994 9:58 am PST
Subject: The Nature of Control
[From Rick Marken (940709.1045)]
Paul George (940709 11:30) and (940708 15:00 EDT) --
I'll try to answer or ask questions about some of Paul's points in anotherpost. But first let me suggest that Paul's basic problem -- the reason for allthe apparently ridiculous statements in Paul's posts -- is a lack ofunderstanding of the nature of the phenomenon of control. It looks like Pauldoes not intend to read the PCT literature or work with the computer demos(which is fine with me since doing so is no guarantee that a person will getthe idea anyway) so before he makes any more (what I'm sure areunintentionally) silly statements, let me explain what we mean by "control" inPCT.
We begin with the observation that, in living organisms (with nervoussystems), efferent neural signals cause actions (variations in muscle tension)that have effects on the environment. We can represent the situation with anequation:
(1) q = f(o)
where o is an action variable (or output) and q is a result of that output.The function f() represents the environmental laws that determine the nature ofthe relationship between output and result. (Of course, any output variableproduces many results; I'm just focusing on one result here -- one that mightbe under control). An example of what is described by equation 1 is lifting abook; q is the position of the book and o is the muscle force(s) that influencethe book's position. f() is the differential equation that transforms muscleforce into book position (taking things like the mass of the book intoaccount). But it is important to note that q can be ANY result of output,including very complex results. For example, q could be "amount read aboutcontrol theory" , a very complex (and variable) result of many outputs andintermediate results of those outputs (like lifting a book); but a result ofoutput, nevertheless.
The next observation is that any result of output is also a result ofnon-organism produced, environmental influences -- disturbances. Assuming thatthese influences add to the influence of the organism (output), the completeequation for the cause of any result of output is:
(2) q = f(o) + g(d)
In this equation, the function g() represents the differential equationsthat relate the disturbance variable(s), d, to the result, q. A disturbance tothe position of the book, for example, is the gravitational force, which isbasically constant; a variable disturbance might be a force on the same bookexerted by someone trying to help you lift the book.
So, at any instant, the state of the result of an organism's outputs (thevalue of q) depends on the state of the organism's outputs as well as the stateof (possibly many) environmental disturbances, d.
We say that a result of output is under control when q remains in aconstant or variable reference state, q*: that is, when q = q*. This onlyhappens when the influence of output on q -- f(o) -- precisely mirrors theinfluence of all other variables on q -- g(d). Letting q* be constant and zero(for simplicity), we say that there is control when:
(3) f(o) = - g(d) (approximately)
Remember that o, d and q are variables -- changing over time. It is,therefore, highly unlikely that f(o) would remain precisely opposed to g(d) forlong simply by chance. Thus, equation (3) is the basis of The Test for control;we apply arbitrary influences to q (we vary g(d)) and look either for preciseopposition from the system (-f(o)) or _no effect_ of variation in g(d) on q. Ifthere is precise opposition to g(d) or no effect of g(d) on q then we say thatq is a _controlled variable_ and it's reference state (which can vary overtime) is q*.
The Test need not be applied quantitatively; we can get hints aboutcontrolled variables by simply "disturbing" them. For example, if we think aperson is controlling the amount they read about control theory, then we candisturb this variable by talking about nothing but politics and religion onCSG-L; a person who wants to read about control theory might protest, questionwhat is going on or try another list (the fact that there are many differentways to influence a controlled variable is why we usually monitor the effect ofg(d) on the controlled variable itself, rather than on f(o), since we rarelyknow all the outputs that might be used to keep a variable undercontrol).
Again, it is important to realize that a controlled variable, q, can be any_variable_ result of an organism's output -- it can be the position of a book(in x,y z space) or one's position on abortion (in political space?). It isalso important to remember that all the results people produce, which meanseverything we call their "doings" -- lifting books, reading about controltheory, going to get an abortion or preventing others from getting one -- allthese results of output are a result of influences of _both_ outputs anddisturbances (equation 2). To the extent that people produce these resultsreliably (produce consistent values of q) then these results must be undercontrol -- that is, outputs must be varying appropriately to counter theinfluence of disturbances to these results (equation 3). This is what we meanwhen we say that all behavior -- all consistently produced results of action --are under control. And we do have evidence -- quantitative evidence -- thatpeople can control simple variables, like the position of a cursor, and "higherlevel" variables, like a sequence or program of numbers that is occurring on acomputer screen.
But the evidence of "higher level" control is all around you; virtuallyevery aspect of the world you live in is a result of people (including you)contriving to produce the results you see rather than the results that wouldhave been produced by the operation of environmental disturbances (even ifthese were combined with carefully pre-programmed outputs,o). Higher levelcontrol is evidenced by the fact that there is a computer sitting on your deskrather than a pile of sand,
Now, don't you feel like starting all over again, Paul ;-).
Best Rick
Date: Sat Jul 09, 1994 8:29 pm PST
Subject: Re: No control without feedback
<[Bill Leach 940709.18:54 EST] >[Bill Powers (940709.0740MDT)]
And thank you, maybe I am learning something after all :-)
>Bill Leach (940708.xxx):
> ... (thinking about that, you actually can control some industrialprocesses without explicit negative feedback control though it tends to make afailure prone system).
As is usual with such discussions, "levels" can be a problem. I can thinkof literally hundreds of actuators that operate with negative feedback control.In some cases there are automatic systems that will detect the error in arather indirect way but in many cases about all the control system perceives isthat there IS something wrong with neither the "knowledge" of what the problemmight be nor the ability to correct the error.
So, actually what I was thinking of is that in some cases simple "loops"are sometimes created where an actuator operates blindly in the sense thatthere is no sensing of either the actuator's operation nor the results of thetask performed by the actuator. This is still called control or at least is apart of control system though the use of the term in those cases is really abit loose (of course if one specifically calls it "open loop control" thenthere would be less confusion).
Indeed, that was exactly that type of "control" that caused the TMIincident. The Power Operated Relief Valve was operated "blind" and failed toclose after the control system opened it to reduce an over pressure condition.The open operation is actually blind also but there is rather obvious negativefeedback in the sense that the controller calling for the valve open conditionwill respond to the resulting pressure drop if the valve does indeedopen.
This is not intended to excuse the operators of the plant for failing tocheck the actual valve condition. It is and always has been common knowledgeamong competent steam plant operators that relief and safety valves should beexpected to leak following operation.
An automobile power steering system is a good example of a hydraulicnegative feedback control system (with mechanical over-ride). A hydraulic metalpunch OTOH is often NOT a closed loop system unless the human operator iscounted as a part of the loop. That is, when the valve opens to admit oil tothe cylinder to drive the punch head it is assumed by design that the punchwill travel the required distance and in a like manner when the cylinder isrelieved (or the withdrawal cylinder is pressurized) it is expected that thepunch head returns.
I don't think that there is necessarily anything wrong with designs thatare open loop as long as the operator that is monitoring or using the equipmentis aware of the fact and its' significance. For really complex systems however,such components are a poor idea.
-bill
Date: Sat Jul 09, 1994 8:56 pm PST
Subject: Re: open-loop "control"
[From Bill Powers (940709.2140 MDT)] Bill Leach (940709.1854 EDT)
> Indeed, that was exactly that type of "control" that caused the TMIincident. The Power Operated Relief Valve was operated "blind" and failed toclose after the control system opened it to reduce an over pressurecondition.
Yes, I've used that example, too. I believe there was a light on thecontrol panel indicating "open" and "closed," but all it indicated was thestate of the switch. It should have been connected to a flow sensor, shouldn'tit? That's what you _really_ want to control. Even connecting it to the valvegate isn't enough: there might be somebody's overalls stuck in the orifice, orthe water supply might have run dry (or, as at TMI), you may think the coolingwater flow is shut off while it in process of draining dry. What you sense iswhat you control.
> A hydraulic metal punch OTOH is often NOT a closed loop system unlessthe human operator is counted as a part of the loop.
That was my point: the loop is _always_ closed, somehow, if there is to becontrol. When you build an open-loop device, it becomes part of the outputfunction of a human being who watches the result to make sure it's what iswanted. Nobody trusts a purely open-loop system, and shouldn't. You don't pressthe actuate button on the hydraulic press without watching what the pressdoes.
I've been watching Nasa Select for the last couple of days, and listeningto the traffic between mission control and the spacecraft personnel. You cansee the results of lessons learned over many years: everything is fed back. Stsbio, Houston. Houston, go ahead. I have a temperature reading from the biorack.Ready to copy. Incubator A reads niner five point 3. Copy niner five pointthree on A. That's affirmative, Houston.
Even when someone asks an astronaut to flip a switch, there are people onthe ground monitoring the effect. A couple of missions ago, this led to tracingdown a communications malfunction to a power supply switch that was flipped tostandby instead of on. This time, it led to figuring out that the biorack powerwent off because an astronaut had caught a switch with his toe (specialproblems of working in zero-g!). And the astronaut didn't just flip it back on.He reported the event, and asked whether or not to turn the power back on.Then, given the go-ahead, he reported that he had switched the power back on,AND that the biorack was again operating. And Mission Control acknowledged:biorack operational again. These people understand exactly what controlmeans.
> I don't think that there is necessarily anything wrong with designsthat are open loop as long as the operator that is monitoring or using theequipment is aware of the fact and its' significance.
But if the operator is monitoring the results, the loop is closed. In atruly open-loop system, NOBODY is monitoring the results. We just don't speakof control when nobody is monitoring the results. We speak of effects orinfluences, not of control (or should). The air temperature at the north poleinfluences the freezing of the sea. But it doesn't control the freezing of thesea. If the temperature drops and the sea doesn't freeze, the temperature won'tdo a durn thing about it.
> For really complex systems however, such components are a pooridea.
For simple ones, too, unless the consequences are unimportant. Remember,for truly open-loop control, nobody gets to see whether the device operated, orwhether its operation had the expected effect. All you really have to do toconvince an engineer that all control is closed-loop is to keep him from seeingwhat happens when he turns on his "open-loop control system."
Best, Bill P.
Date: Sun Jul 10, 1994 1:59 pm PST
Subject: Re: open-loop "control"
<[Bill Leach 940710.17:31 EST] >[Bill Powers (940709.2140MDT)]
Ok, but this is a "perspective" issue to an extent. At some level, probablyno engineered control system is open-loop by design. Portions of it may howeverbe open-loop in the sense that under normal conditions the operation of thatportion is not monitored.
When designing a control system, normally the engineer considers thephysically engineered parts as either open-loop or closed-loop based upon thecontrol system behavior sans observer.
When designing something that is open-loop, a competent engineer will atleast consider how the rest of the system will respond to a control failure inthe open-loop part and how an observer might recognize such failures.
I am not sure that further discussion of this aspect is particularlyimportant to PCT but recognize that viewing all systems that have humanoperators present with a perspective that they are a part of the systemfeedback is a good idea.
I think that much analysis of system failure modes fail to actuallyconsider the operator in such a fashion.
-bill
Date: Sun Jul 10, 1994 4:15 pm PST
Subject: Replies to Paul George
[From Rick Marken (940710.1700)] >Paul George (940709 11:30)--
> (Though I have often thought Skinner was grossly misunderstood)
All Skinnerians feel that Skinner was grossly misunderstood. The fact is,however, that there is nothing to misunderstand about Skinner because he had nounderstanding to miss. Skinner did no modelling so he could change hisfantasies about behavior faster than you could question his theoretical claims.Skinner could no more be misunderstood than could a Talmudic scholar; both knowwhere their reasoning is going and they won't let data and models stand intheir way. For example, I have experimental data that shows that operantbehavior involves selection OF consequences (consistent with PCT), notselection BY consequences (as Skinner argued). Skinnerians have a nifty way ofdealing with this data; rather than showing how a reinforcement model handlesthe data they simply say "you've got it wrong" and blythly press ahead. Quitean interesting "scientific" dialog.
> The question is not "is it right" but rather "does it work".
It's true that people can make things work without knowing the right (best)model of how it works; most human controlling works that way (we have noknowledge, for example, of the nature of f(o), the physical laws that relateour outputs to most controlled consequence of those outputs; nevertheless, weare able to control those consequences; this is just the way control works) butthere are many things that would have been quite impossible to do without theright model (that is, without a detailed understanding of how our outputsinfluence the consequences of those outputs). For example, most high technology(space travel, computer chips) would have been quite unlikely to happen ifpeople did not have the right models of physics and materials.
> A control model is required to capture the _control component_ of thebehavior.
Correct. And that's a very big "component". Now that you have read myprevious post on the nature of control, can you think of ANY behavior that doesnot involve control? That is, can you think of any result of action that is notbrought to an intended level while being protected from the influence ofdisturbances? One behavior I can think of that has no "control component" isthe downward velocity of a person who has jumped off a cliff (with no parachuteor wing); not a very common behavior but a behavior that involves no control(after the jump, anyway). Can you think of any others?
> Further PCT has not (to my knowledge) [shown?] that higher levelbehavior _is_ a "a controlled variable".
Look back to last month's posts for a description of my program controlstudy; I have a copy on HyperCard now. It shows unequivocally that a simpleprogram of events ("if X then Y else Z") can be controlled (maintained againstdisturbance) by non-programmatic outputs.
> I personally suspect it involves control of sets of controlledperceptual variables.
Me too. It's called Hierarchical Control Theory. It's implemented as a realWORKING model. A nice description of a working, hierarchical control model isgiven in my book "Mind Readings", Ch. 6. See how useful reading canbe:-).
> I am not sure that control phenomena is all that exists. There arecomponents that appear reactive.
As Tom said, cause-effect phenomena exist, too. In fact, they are farbetter documented; they are what physicists study.
What appears "reactive" in a living control system, however, is part of thephenomenon of control; it is a reaction to sudden disturbances to a controlledvariable. It looks like the disturbance causes the reaction but that's notquite how it works. The "reactive" component of behavior is disturbance -resistance that keeps the controlled variable at the reference level. Thenature of the "reaction" to a disturbance depends (among other things) on thereference setting for the controlled variable -- which is determined,autonomously, by the organism itself. In my previous post I wrote the equationfor disturbance resistance under the assumption that the reference for thecontrolled variable was fixed at zero. Here I rewrite the equation allowing forthe possibility of a non-zero and variable reference, q*:
f(o) = q* - g(d)
So the nature of the "reaction" (f(o)) of a control system to a disturbancedepends on the reference setting for the controlled variable. What this meansin real life is that the same disturbance will not always produce the samereaction. When someone takes a swing at your face (d) you usually duck (o) inorder to maintain the amount of pain you experience (q) at a reference level ofzero (q*). But if you want to prove how tough you are you might revise yourreference for pain -- and the same disturbance (the swing) results in noreaction (you don't duck) and, therefore, you get hit -- bringing thecontrolled variable (pain) to its new, non-zero reference.
You say:
> When a hand is burned, the arm will jerk back.
Now you can see that this is only true if your reference for pain is zero.There is a great scene in the opening to "Lawrence of Arabia" where Lawrenceholds his hand over a fire to demonstrate the absence of the "jerk reflex" whenhe changes his reference for pain. When asked if it hurts, I think Lawrencesays something like "Of course it does; you just have to not mind". Lawrence ofArabia: control theorist.
> Bill Powers statements recently notwithstanding, a control loop canstop, or at least idle. Computer systems often use something called andinterrupt chain.
The control loops we are talking about go through the environment, wherevariables don't behave like idealized digital circuit elements -- they aredifferentiable functions of time. A programming loop, for example, is not(usually) a control loop, in the control theory sense of control.
> You focus upon the differences, not the commonalities or how theapproaches might be blended into a more powerful whole.
I wrote a paper called "The blind men and the elephant"; it is published inthe CSG journal "Closed Loop". In that paper, I explain the "commonalities"between PCT and S-R, reinforcement and cognitive views of behavior. PCT showsthat these three views of behavior capture three different aspects of control,but each, alone, misses the big picture. PCT shows why behavior has appeared tobe a response to stimulation, selected by consequences or planned output. Infact, behavior is none of these things, though it can look like each under theappropriate circumstances. Behavior is only one thing -- control.
> Ultimately I suspect PCT will meet other approaches somewhere in themiddle,
I wish this were true. Unfortunately, reconciling PCT with other approachesis a bit like trying to reconcile Copernicus and Ptolemy. While it waseventually possible to see why Ptolemy would see things as he did, there was no"compromise" solution (like having the earth be close to the center of theplanetary system -- where Venus is -- instead of three planets out). PCT showswhy the "other approaches" see things as they do; unfortunately, it shows thatwhat these approaches see is not what is. The "elephant" of control is not the"snake" of S-R, the "tree trunk" of reinforcement or the "wall" of cognition.And yet, they control;-)
> When you can build an executable control model showing how one puts ona sweater and/or adjusts a thermostat, and decides between the options, I willbe convinced.
You should only be "convinced" of PCT when you are convinced that behaviorIS control. This is an empirical -- not a theoretical -- question. If you arealready convinced that psychologists know what they are talking about when theytalk about "behavior" then PCT is irrelevant to you.
PCT can only be of serious interest to people who are convinced -- byOBSERVATION AND TEST -- that behavior IS the control of perceptual inputvariables.
If your selection of theories of behavior is just a "beauty context" -- tosee which theory "looks best" then, even if PCT wins your favor, it is only asuperficial victory and you will join the ranks of people (like Carver andScheier, and other useless "adherents" of PCT) who like PCT but don't like whatit is about -- THE PHENOMENON OF CONTROL. The first thing to learn andexperience about PCT is the PHENOMENON that it explains.
I sometimes think that people should be required to test for controlledvariables for a year before being allowed to read anything about the theory ofcontrol. Phenomena first!!
Paul George (940708 15:00 EDT) --
> All paths between nodes are not permissible in the HPCT model either(of course you define internal connection rules).
You seem to have answered your own point in parentheses. Powers explainedhow the HPCT model worked and how it could be implemented on a computer. Albusleaves most of the workings of the model to our imaginations -- which is finebut rather subjective, don't you think?
Powers says:
> And from this model, we conclude that "operant conditioning" is mostlyordinary control behavior, misinterpreted. So it isn't operant conditioningthat works; it's control that works.
Paul:
> Close. Control is _a_ mechanism which _can_ implement operantconditioning, given certain assumptions. How is the control variable andnegative feedback signal set up? (I have a little trouble with 'Unseen Hand'arguments)
In operant conditioning (once the animal has learned how to use the operantapparatus; that is, after "shaping" has occurred) the rate of reinforcement iskept under control; rate of reinforcement is the controlled variable, q. Theso-called "schedule" that relates bar press rate to reinforcement rate is the"feedback" function, f(), that relates outputs to inputs. There are usually nodisturbances introduced in operant experiments but they are quite preciselyresisted when they are introduced; rate of reinforcement is maintained at areference level, q*, despite disturbances or changes in the feedback function.A huge amount of operant data can be explained precisely by assuming thatorganisms control a perception of the rate of reinforcement by varying theiroutput (bar press rate) appropriately to compensate for the changingcircumstances (schedules mainly). So reinforcement doesn't "control" behavior (now that you know what control means, wouldn't you find it amazing if it did?);behavior (the outputs of the organism) control reinforcement.
> Note: I don't really intend to have a lengthy discussion of the propertranslation of OC to a control model.
That was clear from your brief description of what would be required totranslate OC to a control model. I suggest that you read the description of acontrol model of operant conditioning starting on p. 67 of Powers "LivingControl Systems" to see a proper translation of OC to a control model.
Best Rick
Date: Mon Jul 11, 1994 12:10 pm PST
Subject: Re: Replies to Paul George
[From Paul George (940711 ????)] >[Rick Marken (940710.1700)]
> All Skinnerians feel that Skinner was grossly misunderstood.
I'm not a Skinnerian. But I do think some of the ideas in B&D andWalden II were useful. There is wheat among the chaff. And you malign Talmudicscholars. OTOH philosophy is like that.
> And that's a very big "component". Now that you have read my previouspost on the nature of control, can you think of ANY behavior that does notinvolve control?
I agree, and said so earlier. Never suggested that any behavior (above thereflex level) didn't involve control. Reflex can be overridden, but I am notyet convinced that reactive mechanisms don't exist in nature. As I saidelsewhere, the term behavior covers a very wide semantic range. It does notalways refer to the same thing.
> What appears "reactive" in a living control system, however, is partof the phenomenon of control; it is a reaction to sudden disturbances to acontrolled variable.
Forgive a question probably answered in the literature. How do you knowthis for a fact? Can you empirically show each and every controlled variable,or just that they can exist? How do you identify a variable in tissue? Reactivebehavior can be _modeled_ using control, but it does not necessarily followthat the map is the territory. Note that I do not disagree that behavior can bebased upon perceptual control, just that it must be. That has not beenproven.
> When someone takes a swing at your face (d) you usually duck (o) inorder to maintain the amount of pain you experience (q) at a reference level ofzero (q*).
This is the level where your model fails to convince me. There may be agestalt phenomena, but the concept of ducking to keep a pain variable at 0 is abit of a stretch. I don't disagree that the behavior involves control, I justhave trouble with the described mechanism. The model requires considerableelaboration.
> I wrote a paper called "The blind men and the elephant"; it ispublished in the CSG journal "Closed Loop".
Any way to get an electronic copy? I can't get into anywhere the journal iscarried.
> In operant conditioning (once the animal has learned how to use theoperant apparatus; that is, after "shaping" has occurred) the rate ofreinforcement is kept under control; rate of reinforcement is the controlledvariable, q.
Sorry, I was refering to the likes of dog training, not the controlledexperiments used to document the principle. Again, I am trying to apply the PCTmechanisms to reality (not to imply that they don't apply).I am focusing ondesign of control hierarchies to reflect actual behavior, i.e what would be thecontrolled variables, inputs, outputs, and control hierarchy. We should be ableto define a HPCT system to cover known any behavior, whether or not thetechnology exists to actually implement it.
Date: Mon Jul 11, 1994 2:30 pm PST
Subject: Re: PCT models
[Paul George 940711 11:30 EDT]
>[From Tom Bourbon 940708.1653]
> Should I wait until Monday, so I can cool down a bit? Nah! Would youlet us in on what you found the most amusing? And would you maybe let us in onyour personal assessment of Albus, now that we know you weren't playingstraight with us for the past week and a half? Whatever your assessment may be,I guess we don't need to wait to see any implementations of a genuine Albusmodel by you, do we?
Similar sentiments from [From Dag Forssell (940708 1800)]
Cool Down allready! Mighty thin skin for senior (I presume) researchers.Dry humor seems to fall on deaf eyes. I guess I have to make greater use ofsmilies, and be very careful about phrasing and images.
I wasn't playing with you. I suggested the article since it seemed toprovide a sounding board for PCT. It _was_ a sincere question. I got muchreaction and discussion that revealed much about PCT, which was my intent. As Isaid repeatedly, I didn't consider the Albus article to be a seminal work, justinteresting. We are designing a control system using some of Albus' concepts,but that isn't the kind of implementation you are interested in. I am anengineer, not a scientist, even if that's not the way I planned it many yearsago.
FYI I find discussion amusing, and a number of the comments and examplesthat have been made. Amusement is not the same as mockery. I will read theliterature as soon as I can get my paws on it. So far I've had no luck inlibraries, and I don't have access to a university that has a strong psychologyor neurological department.
> I have a long-standing interest in Skinner and in the often-repeatedclaim that he was grossly misunderstood. Could you give a few specificexamples of what you mean?
It has been a number of years since I read Skinner's books. Other schoolsoften argued he felt that there was no such thing as intelligence orpersonality, just conditioned response. My reading was that he felt that onecould not determine the existence of 'soul' and such empirically and so oneshould stick to what could be observed. He felt that OC was a useful techniquefor training humans to get more 'civilized' behavior. And he thought thattreating people as anything more or less than animals was at best hubris. Thisenraged the Humanists in particular. I always thought they were saying much thesame thing from different angles, but were too busy attacking one another toallow communication. Of course professorships and tenure were on theline.
> Further PCT has not (to my knowledge) that higher level behavior _is_a "a controlled variable".
> Something is missing from your post.
Sorry, too much cutting and pasting. Should have read "demonstrated" as inexperimentally. PCT makes the assertion that 'all behavior is control', where Iwould suggest that something like "all behavior involves control" is likelymore accurate.
> I would still like to know how you might program and run the model forthis kind of performance. Albus did not provide that level of detail in his"outline." And I would like to know if you think such a model would be arepresentation of how living systems function.
I suspect it describes how organizations and high level 'reasoning' typebehavior works. I doubt it goes all the way down. At some point the concepts ofvalue judgement and world model atrophy into a reference variable value. I alsosuspect that something like a PCT control system is involved in the internals.I can think of no way to prove such a model except by building a trueartificial intelligence, and that would only prove that the model _can_ work,not that living systems are in fact is constructed that way.
> That happens to be one of the models I will demonstrate at theupcoming meeting of the CSG, in Colorado.
I would be interested in the particulars. The hints you and Bill Powershave given in other posts made it appear to be at a bit lower level than I wastalking about. Part of the problem may be that the word 'behavior' is used torefer to everything from reflex to painting the Sistine Chapel.
> Oh, and by the way. Would you tell us about some of the people whohave already built the models for putting on a sweater and adjusting athermostat?
My point exactly. No one has yet built working models of high levelbehavior without gross oversimplification. When we do we will have created anartificial lifeform.
>[Bill Leach 940708.21:38 EST(EDT)]
Thanks for a lot cooler post.
> The reason that this is actually dangerous to the world is that, ifthe basic PCT concept is right -- and there is compelling evidence that thebasic concept is indeed right --, then these behaviouralists are treatingsymptoms and not causes.
> Behavioral scientists are not rigorous. They basically all run aroundsaying that human behavior is too complex to be able to prove a theory. Havingaccepted that idea, they then refuse to either examine hard science data orhave such principle based tests applied to their own theories.
I concur, and that I why I find CSG and PCT so interesting. You haveexperimentally demonstrated the underpinnings, and generalizing the theoryallows insight into higher level behaviors and clinical applications. Thereremains however a lot of work in the middle before we can say "all behavior iscontrol" (or more accurately is the result of controlling perception). {Notethe 'we'. I think you are on the right track, if occasionally locked into theworm's eye view}
>> It is kind of like the difference between a 'top down' and 'bottomup' approach to analysis. Both have their pitfalls.
> This sounds like a valid argument but it is not (in my opinionanyway). "Top down" and "bottom up" both deal with exactly the same knowledgeand exactly the same understanding of operational details.
Using analogies is always risky. Actually In my area (software engineering)different levels of a model contain very different knowledge. Understanding ofthe operational details also varies widely (a severe problem). One of the mostinteresting things about working with OJT trained developers and users is thewildly varying assumptions they make about how things actually work in anapplication and how the hardware functions. Yes, "Someone somewhere didactually fully understand the details of hardware operation". But developer'soften do not know the details, nor do they care.
I still think the analogy is valid. A top down approach takes the bigpicture and attempts to decompose it 'logically' (variously defined) intocomponents. The bottom up attempts to create structures from existingcomponents in order to meet goals. The two approaches can produce verydifferent architectures. However, we have found that a better technique is towork from both directions meeting a third effort expanding from the middle (inreality you jump around).
The point is that different abstractions are useful at different levels ofdetail, and from different points of view. Asserting that only one theory isvalid is dangerous, though occasionally right. I often find the parable of theblind men and the elephant enlightening.
Take care Paul George
Date: Mon Jul 11, 1994 3:48 pm PST
Subject: Re: The Nature of Control
[Paul George 940711 12:00] >[Marken (940709.1045)] (940710.1700)]
Thanks for the primers.
A number of you seem to be working under the misapprehension that I don'tunderstand control from an engineering standpoint and that I disagree with PCT.I don't. I am not attacking you. I have no axes to grind, nor do I belong toany competing school of thought. Remember, _I_ am not a behavioral psychologistand have no vested interest in your being wrong.
Is there some reason you feel that PCT is not understandable? I know thatyou are apparently used to being misinterpreted by others in your field. Yourexamples make perfect sense to me and are rarely surprising. The higher levelconcepts flow logically from the basics. They largely match my own thinking.While I have more reading to do, this group has brought out the core of PCTquite well. I do have some questions about _how_ it scales up, and themechanisms used to assemble and modify the controlled variables and controlhierarchy. These are not defects, just areas for future work. I also tend toquestion phrasing that appears to 'fluff the wares'.
What you usually seem to read in my posts is rarely the point I am tryingto make. Perhaps I use analogies with which you are unfamiliar, or terms thatproduce a reflex response. I try to feed back your examples from a differentangle to produce discussion (I tend to play devil's advocate). The responses(from my point of view) seem to home in on side details that I didn't botherfleshing out or re-iterate points that were eloquently described earlier. Wedon't seem to be focusing on the same issues. This indicates that I am failingto communicate. But please be open to the idea that you actually transmit PCTconcepts very clearly. Don't presume misunderstanding.
Date: Mon Jul 11, 1994 3:49 pm PST
Subject: Re: Testing models
[Paul George 940711.15:00] <[Bill Leach 940708.23:01EST(EDT)]
>> A PCT control node must ignore all inputs that do not affect thecontrolled variable, or more accurately never receives them.
> Interesting assertion. Why does this have to be true?
Perhaps I should clarify. This relates to the thread on open loop control.The universe provides a nearly infinite variety of inputs, or sense-ablethings. The body has a limited set of sensors. Their signals must routed to acontrol structure that interprets them in terms of some hierarchy of controlvariables. Sensory input is distinguishable from perception. A HPCT network hassome set of controlled variables at a given time. Similarly the nodes'generate' a given set of possible actions to attempt to control thosevariables. Part of the universe of possible perceptions are thus 'open loop' ata given time, as they are not being controlled. Many of the results of ouractions are not perceived, as are the actions themselves.
One of the things I do not yet understand in PCT is how an organism'decides' to control a perception and then discovers the actions that willaffect it. It appears to be a bit of a chicken and egg problem. It even worsewhen a Higher node must set up a control relationship with lower nodes whichactually do the sensing and acting. {I did read Bill Powers post on howorganisms develop, and found it persuasive, if not complete}If just randomchance was at work we would expect the control solutions to vary radically fromindividual to individual. OTOH we observe that people and other organisms seemto share control variables and reference values.
Best Paul
Date: Mon Jul 11, 1994 6:14 pm PST
Subject: Re: Testing models
[Avery.Andrews 940712.1022] (Paul George 40711.1500)
> One of the things I do not yet understand in PCT is how an organism'decides' to control a perception and then discovers the actions that willaffect it. It appears to be a bit of a chicken and egg problem.
It's either hard-wired, or stumbled across by reorganization. Cockroacheshave some hard-wired circuitry that figures out from what direction a puff ofair is directed, and sets a `direction of turn' parameter appropriately. On theother hand, people can alter the signs of feedback loops involved forcontrolling their behavior - when a sign is wrong, you get a very distinctiverunaway effect. In fact, Bill Powers and Rick Marken discover that peoplespontaneously reverse signs during even highly practiced tracking tasks (in oneof the Mind Reading articles, I believe), indicating, I think, a constant lowlevel of reorganization going on in the background.
In between human flexibility and cockroach rigidity, there is a broad rangeof variation. In one of the chapters of the Gallistel's `Behavior, a NewSynthesis', there is some discussion of the apparent fact that birds that usetheir feet to manipulate food can be taught to operate food-dispensingmachinery with their feet, birds that use their beaks for this purpose can betaught to operate food-dispensers with their beaks, but not vice-versa. It issaid to be impossible to condition a pigeon to operate a food-dispenser byhitting a lever with its wings. This suggests that there is some hard-wiring ofthe reorganization system to the effect that an unsatisfied reference level forsome physiological variable connected to food intake causes reorganization in aspecific area of the motor-control system, varying with the species.
The reptile-amphibian divide is also important here: if you cross thenerves operating an amphibian's limbs, the central activation patterns won'tchange, and they will for example scuttle backwards from food instead of towardit (I recall this from an old scientific american article - there must be somemore recent work on this). Lizards on the other hand can, I think I remember,reorganize their way out of this inappropriate structure, as of course canmammals.
So the basic idea is that unsatisfied higher-level reference signals causelower-level systems to reorganize on an essentially random basis (althoughthere may be fixed constraints on where the reorganization caused by aparticular kind of error happens), & reorganization continues until thehigher-level errors stop. Obviously, there is a huge amount of detail left tobe filled in here, and I would very much like to see a demo in which asimulated creature learned to do something interesting by reorganization (infact, I'd like to write one, but haven't worked up to the point where thatwould be a sensible project to take on).
Avery.Andrews@anu.edu.au
Date: Mon Jul 11, 1994 7:32 pm PST
Subject: Re: Testing models
<[Bill Leach 940711.21:54 EST(EDT)] >[Paul George940711.15:00]
> Perhaps I should clarify. This relates to the thread on open loopcontrol. The universe provides a nearly infinite variety of inputs, orsense-able things.The body has a limited set of sensors. Their signals mustrouted to a control structure that interprets them in terms of some hierarchyof control variables. Sensory input is distinguishable from perception. A HPCTnetwork has some set of controlled variables at a given time. Similarly thenodes 'generate' a given set of possible actions to attempt to control thosevariables. Part of the universe of possible perceptions are thus 'open loop' ata given time, as they are not being controlled.
Speaking of precision in use of terms... Is the use of "universe" the samein both instances?
One of the problems in understanding PCT, I suppose, is that perceptions donot necessarily have to be under control. That is not what PCT says, only thatbehavior is the result of controlling perceptions.
In this area, I may be "stepping out" beyond my expertise but will riskit:
It appears that all biological control systems are "unidirectional". Bythat I mean that that reference signal can only run from zero to somemagnitude. Thus, in any given control loop there is no such thing as reversingthe sign of control.
From a physical standpoint, this is actually rather obvious. If you want toturn your head to the left and then to the right, you don't reduce the "turn tothe left reference until it changes sign" but rather actually evoke a differentcontrol system to make the change (portions of the systems may be common butnot the portion that actually effects the rotation). Even in a "simple" examplesuch as this one, there are many perceptions actually being controlled toaccomplish the "single goal."
Many control loops must have their references set to zero -- there is nohigher control loop that requires their operation.
> Many of the results of our actions are not perceived, as are theactions themselves.
I am not sure what the significance of this statement is suppose tobe.
> If just random chance was at work we would expect the controlsolutions to vary radically from individual to individual. OTOH we observe thatpeople and other organisms seem to share control variables and referencevalues.
This is again a reason to call for reading the literature. I will remark inone way though... have you studied the behavior of an infant much?
-bill
Date: Mon Jul 11, 1994 7:46 pm PST
Subject: Re: PCT models
<[Bill Leach 940711.21:27 EST] >[Paul George 940711 11:30EDT]
> I think you are on the right track, if occasionally locked into theworm's eye view}
I think that I have addressed some of the source of this "view of PCT". Ashas been suggested, you really have to read some of the PCT literature to seejust how much thought has gone into the "challenges" that you raise. In manycases, suggestions that have been made for changes to the PCT paradigm can beshown to be impossible requirement for living systems (using real data acquiredfrom various sources).
> I still think the analogy is valid. A top down approach takes the bigpicture and attempts to decompose it 'logically' (variously defined) intocomponents. The bottom up attempts to create structures from existingcomponents in order to meet goals. The two approaches can produce verydifferent architectures. However, we have found that a better technique is towork from both directions meeting a third effort expanding from the middle (inreality you jump around).
At issue here in this minor disagreement is that while both methods work,they are neither inconsistent with reality (when they both work). If whenemploying either method, one is using an incorrect fundamental concept set, theeffort will fail.
The idea that behavior is ONLY the result of controlling for perceptions israther unique view in behavioral science (yes disturbances affect behavior butONLY because the disturbances cause changes in the perception).
HPCT claims (rather convincingly in my mind) that the paradigm of negativefeedback control works equally well for the thought processes that I amexperiencing while I write this to you. There are one hell of a lot of manyears of thought and experiment involved in coming to that conclusion as wellas many well thought out challenges to the idea.
-bill
Date: Mon Jul 11, 1994 9:02 pm PST
Subject: Re: The Nature of Control
<[Bill Leach 940711.21:42 EST(EDT)] >[Paul George 94071112:00]
> I do have some questions about _how_ it scales up, and the mechanismsused to assemble and modify the controlled variables and control hierarchy.These are not defects, just areas for future work. I also tend to questionphrasing that appears to 'fluff the wares'.
While PCT might well be the "richest" field as far as new work isconcerned, I think you will be more than mildly surprised as how much of whatyou have been talking about is anything but new to PCT.
> Don't presume misunderstanding.
I don't believe that anyone has presumed misunderstanding. The responses toyou postings have been quite genuine and honest.
To give you a little example of just how a PCTer might have to "beatsomeone over the head" to get them to understand the real issue involved in atopic, just go back a month of two in the CSG-L archives and take a look at myexchanges on "Society" with the net. Yes, sometimes I was undoubtedlymisunderstood but I actually always was missing a point (it is just that itoften was not the point that others thought that I misunderstood -- it wassomething else).
These guys tend toward extremes of precision, particularly with fundamentalPCT/HPCT concepts and especially with a "new comer". I used to think myselfthat they got a bit "carried away" but now have come to appreciate that it isprecisely this dogged insistence on getting down to specifics and locking ontoexact meanings for terms that makes it possible for PCT to deal with thecomplexity of behavior in a consistent, reliable, reproducible manner.
Speaking for myself, it really can be a bit exasperating for someone new toexperience but worth the trouble.
-bill
Date: Mon Jul 11, 1994 9:34 pm PST
Subject: Replies to Paul
[From Rick Marken (940711.2200)] Paul George (940711.?)
> I am not a Skinnerian. But I do think some of the ideas in B&D andWalden II were useful.
Not surprising; Skinner was heavy into "using" behavior to achieve his ends-- ie. controlling behavior. His ideas may be "useful" but PCT shows that theyare also a recipe for intra- and inter-person conflict.
> There is wheat among the chaff.
Could you give just one example of Skinnerian wheat?
> Never suggested that any behavior (above the reflex level) didn'tinvolve control.
Ok. But you'll have to throw the reflex level in, too, because reflexes arecontrol processes themselves.
> Can you empirically show each and every controlled variable, or justthat they can exist?
In principle, all controlled variables can be empirically shown to existusing The Test for the Controlled Variable. In practice, we just try toidentify the variables controlled in examples of behavior that happen to be ofinterest at the moment -- like pointing at a target or navigating up a chemicalgradient. The point of PCT is not to discover every variable controlled byevery organism. The point is to show that organisms do control variables. Onceyou see that a variable, like reinforcement rate, is under control, yourconcept of "what is going on" with a particular behavior, like operantconditioning, changes completely. In operant conditioning, for example, yourealize that you don't need a model of reinforcement because there is no suchthing as reinforcement (events that strengthen responses); you need a model ofwhat is actually occurring in operant conditioning -- control (by the organism,not by the environment).
> Reactive behavior can be _modeled_ using control, but it does notnecessarily follow that the map is the territory.
"Reactive behavior" is simply an aspect of the phenomenon of control; whencontrol is occurring, there is simply no way to model the reactive behaviorassociated with it except with a control model. If the reactive behavior is notassociated with control (if, for example, it is like the "falling" behaviorthat is a reaction to gravity) then a non-control model (such as Newton's laws)will work fine.
Me:
> When someone takes a swing at your face (d) you usually duck (o) inorder to maintain the amount of pain you experience (q) at a reference level ofzero (q*).
You:
> This is the level where your model fails to convince me.
I wasn't trying to convince you about the model of control; I was trying toconvince you of the existence of the phenomenon of control as it is manifestedin the behavior of a living system.
> There may be a gestalt phenomena, but the concept of ducking to keep apain variable at 0 is a bit of a stretch.
Why? What is stretched? What does "gestalt" have to do with anything? If,every time you swing (always from a different direction with a different speed), I move my head appropriately so as to avoid the swing (and feel no pain), Iam controlling pain; disturbances (swings) do not have their expected effect(producing pain); their effect is systematically resisted; this is control;where's the stretch? A perceptual variable (pain) is clearly being controlled;the output (head movement) is always exactly what is required to maintain avariable (pain) in a particular state. Control is indicated by the fact thatthe output is always exactly what is needed to compensate for the effect ofdisturbance, keeping a consequence of head movement (degree of pain) in areference state (zero). This phenomenon (controlling pain -- keeping it atzero) could not be produced by a device that produces outputs in response todisturbance inputs; it can only be produced by a closed loop control systemthat is controlling a perceptual variable -- pain in this case.
> I don't disagree that the behavior involves control, I just havetrouble with the described mechanism. The model requires considerableelaboration.
If you agree that the behavior [moving the head in reaction to a swing]involves control then I suggest we put your claim to the test; let's see theelaborations that you think are needed to make the model work.
> Any way to get an electronic copy [of "The blind men and theelephant"]?
It might be on the biome server. I actually don't know. It's one of thepapers that I was never able to get published in a "real" psychology journal.I guess they were less interested in "reconciling" than I was.
> Sorry, I was refering to the likes of dog training, not the controlledexperiments used to document the principle.
Same thing applies; the dog is controlling some variable-- attention, food,petting -- or trying to. During training, the dog is randomly varying it'soutputs until the value of the controlled variable is close to its reference.If rover has to lift his paw when you say "shake" or bark when you say "speak"in order to get his controlled variables where he wants them, then that's whathe does.
Best, Rick
Date: Tue Jul 12, 1994 8:29 am PST
Subject: Control
[From Rick Marken (940712.0830)]
Paul George (940711 12:00) re my "Nature of control" post
> A number of you seem to be working under the misapprehension that Idon't understand control from an engineering standpoint and that I disagreewith PCT.
Many of the things you said suggested a lack of understanding of the natureof control as it is manifested in the behavior of living systems. Knowledge ofcontrol engineering is no guarantee (we have found, to our great dismay) thatone understands the controlling done by living systems . In fact, thecontrolling done by living systems is typically very hard to see. For example,disturbances and the compensatory responses to them are often invisible;behaviors (like lifting a book) just seem to happen. You rarely notice thedifferent forces (actions) that are and MUST be used each time to produceconsistent results. In fact, the only person who actually noticed this "simple"fact about behavior is W. T. Powers. That little observation --that what wecall "behavior" consists of controlled results of action-- combined with therealization that what must be controlled is a perceptual representation ofcontrolled results, is the basis for a monumental revolution in thinking about,studying and dealing with living systems, one for which most life scientists(and control engineers, for that matter) are clearly not ready. This may seemlike an extreme claim but it helps me understand why there are only about fivepeople in the world doing PCT research.
Best Rick
Date: Tue Jul 12, 1994 9:29 am PST
Subject: Re: Testing models
[Paul George 940712 10:30] <[Bill Leach 940711.21:54EST(EDT)]
> Speaking of precision in use of terms... Is the use of "universe" thesame in both instances?
In the mathematical sense (i.e. set theory) yes. The first 'universe' isthe objective world, if any. The second is the set of possible perceptionsbased upon available sensory inputs.
> One of the problems in understanding PCT, I suppose, is thatperceptions do not necessarily have to be under control.
No problem. It is an obvious conclusion, if not the essence. We are notcontrolling the universe or behavior, just _a set_ of perceptions vs _a set_ ofreference values.
>> Many of the results of our actions are not perceived, as are theactions themselves.
> I am not sure what the significance of this statement is supposeto be.
Just a re-iteration of the PCT view that behavior is not directlyperceived, not is the effect of the behavior on the universe. We perceivenothing that we are not controlling for.
> This is again a reason to call for reading the literature. I willremark in one way though... have you studied the behavior of an infantmuch?
Yup. Years of Kittens, a three year old, and one on the way. As Averypointed out in his response, some things are hard wired. Some things areassembled. There are apparently rules or heuristics that guide development.Exploration is not precisely a drunkards walk search. Tuning of control systemshas random elements, put still appears goal directed.
Date: Tue Jul 12, 1994 9:45 am PST
Subject: Re: Replies to Paul and Jeff
[Paul George 940712 10:30] >[Rick Marken (940711.2200)]
> Could you give just one example of Skinnerian wheat?
How about what you cited. Much of his work suggested that people could beguided away from conflicting behavior (as externally perceived) by setting upreinforcement structure that inclined their behavioral patterns towards'better' patterns. Behavior (i.e. my perceptions of what you do) may bemodified by altering the inputs to you. This causes your perceptions to bedisturbed and perhaps adjust themselves so that my inputs resulting from youroutputs are closer to my reference values. There is of course a kind ofHeisenburg effect involved. Skinner didn't put it this way, but the concept isthere. If you wish its 'truth' can be demonstrated by the fact that we cansuggest a control model for the observed results. In my experience childraising works, and I am certainly trying to modify my son's behavior. Trulytrying to control another in all things is very difficult and will causeconflicts, unless I can understand your control structure and control yourinputs (e.g. brainwashing or indoctrination).
> In principle, all controlled variables can be empirically shown toexist using The Test for the Controlled Variable.
Eh, I'll hold off major comment until I can get hold of the literature.However, it is not clear to me that because I can model the system using acontrol variable (or network thereof) that it follows that it is the onlymechanism that can work. {note that I am impressed with the resolve shown touse only this simple negative feedback mechanism to model all behavior; may yousucceed} This appears similar to saying that since I can model control using adigital system that all life must be digital. Different functions can producethe same behavior. They are equivalent, not equal.
> What does "gestalt" have to do with anything?
A set of control nodes containing controlled variables produce the observedbehavior, i.e it is the behavior of the system, not the parts. There _need_ notbe a master variable. Distributed control systems do not necessarily have tohave 'master' nodes.
> If you agree that the behavior [moving the head in reaction to aswing] involves control then I suggest we put your claim to the test; let's seethe elaborations that you think are needed to make the model work.
The example you use is familiar, being a martial artist. At a high levelthere is an observed 'desire' to avoid harm. This doubtless involves avoidingpain. To avoid getting hit I can duck, sidestep, parry, or strike firstinterrupting the punch. Some network of control determines that a punch iscoming, is likely to hit, and might cause harm. Some network of control'selects' the appropriate 'pattern' of behavior, probably by activating othercontrol loops 'programmed by long training. I am likely controlling many highlevel variables such as status (getting knocked down is losing face), form(pride in technique), principle of minimum force, etc. If I was justcontrolling for pain I would duck when anything hurt.
I hope you see that what I mean by elaboration of the model is anelaboration of the set of controlled variables, reference values, and thecontrol network; not the model of perceptual control. Note that different setscould produce the same behavior. Much of martial arts training is simplifyingor tuning the control structure to speed response and effectiveness. Thesimplest structure is not usually the one that forms 'naturally' orinitially.
I don't argue that the phenomenon of control, particularly "produced by aclosed loop control system that is controlling a perceptual variable " is mostapplicable. At worst I would change "a perceptual variable" to "a set ofperceptual variables".
Date: Tue Jul 12, 1994 11:17 am PST
Subject: Testing models
[John Anderson (940712.1330)] > [Paul George 940712 10:30]
>> One of the problems in understanding PCT, I suppose, is thatperceptions do not necessarily have to be under control.
> No problem. It is an obvious conclusion, if not the essence. We arenot controlling the universe or behavior, just _a set_ of perceptions vs _aset_ of reference values.
>>> Many of the results of our actions are not perceived, as arethe actions themselves. I am not sure what the significance of this statementis suppose to be.
> Just a re-iteration of the PCT view that behavior is not directlyperceived, not is the effect of the behavior on the universe. We perceivenothing that we are not controlling for.
Paul, this seems contradictory. First, you appear to say that you
agree with Bill Leach's statement that perceptions do not have to
be under control, and then you turn around and say "We perceive
nothing that we are not controlling for". Or do I misunderstand?
John E. Anderson Beckman Neuroscience Center
Date: Tue Jul 12, 1994 11:18 am PST
Subject: Re: top-down, bottom-up, etc.
[From Bill Powers (940712.0745 MDT)] Bill Leach (940711.2154)
> One of the problems in understanding PCT, I suppose, is thatperceptions do not necessarily have to be under control. That is not what PCTsays, only that behavior is the result of controlling perceptions.
Still not quite the way to put it. Behavior is the _means_ of controllingperceptions. As you say, not all of a particular organism's perceptions aresystematically affected by that organism's behavior, so not all perceptions areunder control at a given moment. But all behavior is produced by that organismas a means of controlling _some_ perception. Action is generated only in orderto control some perception.
Actions have an array of effects on the world. To an outside observer it isnot always obvious what the controlled effect is. To find it when in doubt, youapply the Test for the controlled variable, which will distinguish unintendedeffects of actions from intended ones.
Nice post.
-----------
Paul George (940711.1130)--
> Dry humor seems to fall on deaf eyes. I guess I have to make greateruse of smilies, and be very careful about phrasing and images.
Dry humor and condescending amusement are not the same thing. Rather thanchoosing more careful phrasing and images, it might be more conducive toavoiding flames if you spoke whereof you know rather than whereof you guessabout PCT.
> We are designing a control system using some of Albus' concepts, butthat isn't the kind of implementation you are interested in.
I think you will find that we are very interested in any actualimplementation of any model of behavior. One of our complaints about criticsfrom other fields is that they dispute the PCT model without offering eitherany criticisms of it or any substitutes for it. If you have an alternativemodel, by all means lay it out to us.
RE: Skinner
> He felt that OC was a useful technique for training humans to getmore 'civilized' behavior. And he thought that treating people as anythingmore or less than animals was at best hubris.
The most interesting aspect of B. F. Skinner as a theorist is that herecommended using one theory but lived by another. He had a goal of helpingpeople toward more civilized behavior, a goal which existed inside himself. Hetried to act in a way that would move his perceptions of how people do behavecloser to the way he felt they should behave. He felt that hubris was not adesirable character trait, so he adjusted his actions, and recommended thatothers adjust theirs, so as to change from being hubristic to being less so. Hehad a low reference level for hubris, obviously. In the laboratory, herecommended getting in mind a firm conception of the behavior that anexperimenter wanted to see an animal performing, and then issuing rewardswhenever the actual behavior was observed (i.e., perceived) to change a littlein the direction toward the desired behavior. He emphasized that this procedurewas not just a rote sequence of actions, but had to be adjusted on the basis ofwhat the animal was doing at every moment, in relation to the desired behavior.This perceptual control process he called shaping.
So B. F. SKinner himself was a negative feedback control system of the typedescribed under PCT, and recommended that other experimenters behave that way,too (as if they could do otherwise). However, the experimental animals andpeople to whom he applied his methods were NOT negative feedback controlsystems. Instead, they were systems that blindly emitted behavior into theenvironment under the control of discriminating stimuli and the consequences ofthat behavior, which he labelled reinforcers if they were successful inmaintaining the behavior. Skinner would vehemently have denied that he hadgoals "in mind", of course, saying "the environment made me say that."
RE: top-down, bottom-up
> A top down approach takes the big picture and attempts to decomposeit 'logically' (variously defined) into components. The bottom up attempts tocreate structures from existing components in order to meet goals.
The top-down approach is a snare and a delusion. It sounds fine when putinto words, because we have lots of abracadabra words to get us past the tightspots. "Decomposing" the big picture into components is actually impossible ifdone strictly from the top down. What you actually do is set up a trial set ofcomponents, see if they add up to the big picture you have in mind, and alterthe components, if they don't add up, until the correct big picture isperceived. So all the judgements are being made in the bottom-up direction, notthe top-down direction.
"Decompose" is one of those dormitive-principle words. It means whateverprocess will create a set of components that, viewed together, add up to avalid instance of the thing to be decomposed.
Suppose the Big Picture is "Wash the car." This task, according to top-downphilosophers, would be analyzed into subtasks such as "get car out of garage,""find hose, bucket, soap, and sponge", and "apply washing method tocar."
The problem is that this "decomposition" assumes that you already know whatthe parts will be. There is nothing in the Big Picture of washing the car thattells you what you need to do in order to achieve that result. The way youdecompose the big picture depends on what you already perceive about the actualsituation at the time the main task is to be carried out. If the car is in thestreet, you can't get it out of the garage. If the sponge can't be locatedyou'll have to find a rag. And if circumstances are just right, you won't haveto apply the washing method; your son might do it for you, or it mightrain.
There are many different sets of tasks that would add up to the Big Pictureof washing the car. You can't get from a statement of the top-level task to adescription of a particular unique set of methods that happens to be workableat the time you decide to do the task. All you can really do is set theperceptual goal of seeing the top-level task being done, and if you don'talready perceive it being done, send signals to more detailed systems whichwill try to supply the perceptions needed to match the top goal with an actualperception.
As the lower systems work on producing candidate possibilities, the toplevel is continually putting the perceived elements together into some BigPicture, comparing that result with the desired Big Picture, and saying "Keepgoing, keep going, keep going, STOP, that will do."
It's the perceptual side, not the output side, that determines whether agiven set of subtasks adds up to accomplishment of the higher task."Decomposition" as a top-down process is a myth.
> Asserting that only one theory is valid is dangerous, thoughoccasionally right. I often find the parable of the blind men and the elephantenlightening.
Ah, then I take it you have obtained Rick Marken's book, "Mind readings,"and have seen his article called "The blind men and the elephant," in which heshows how S-R theory and top-down cognitive theory capture different butlimited aspects of the organism described by PCT.
Best to all, Bill P.
Date: Tue Jul 12, 1994 11:44 am PST
Subject: Comments from Mary
[from Mary Powers 940712]
In reply to Paul George "RE: Replies to Paul George"
Rick said: "can you think of ANY behavior that does not involvecontrol?"
You said: "I agree" - and then excepted reflexes and reactive behavior. Andwent on to ask : "How do you know this for a fact?" And continued, havingdisagreed with the model: "note that I do not disagree that behavior can bebased on perceptual control, just that it must be. That has not beenproven".
It's a little difficult to know where you stand ;-)
Where PCT stands is here: we have a model of the organization of livingsystems. We have applied this model, or hypothesis if you will, as a means ofanalyzing a variety of facts scattered through the literature of human andanimal behavior (bacterial chemotaxis, enzyme systems, neural physiology,reflexes, operant conditioning, tracking tasks, movement of people in crowds,and so on. The same model (which is far more elaborated and sophisticated thanyou have so far troubled yourself to find out) applies in all cases. Using thePCT model, various of these phenomena can also be simulated with high accuracy.Other models, such as S-R and planned-output, cannot be simulated - or if theyare, do not produce the claimed results. In many cases, there are no models atall, just great steaming piles of facts.
The purpose of a model is to organize facts. Proving, or testing, a model,of course is necessary. There does come a point, however, where one canlegitimately project one's model into untested territory with a highexpectation that it will continue to hold. Aren't you pretty confident that ifyou let go of the next thing you hold in your hand, it will fall to thefloor?
This is not to say that it's time for PCT to quit testing - huge amounts ofexperiments need to be done. And they will be done, poking and probing at gapsand flaws. But, if you don't mind, PCTers will conduct those tests withoutconcluding in advance that the model doesn't apply. Simply saying off the topof your head that it does not is far too superficial. It's like refusing tolook through a telescope because (gasp) you might see the moons ofJupiter.
While no one that we know of has trained dogs _consciously_ using PCTprinciples, that would certainly be a situation "in the wild" where theprinciples apply - however you go about training a dog. You exploit the dog'scontrolled variables, either giving him things he likes when he's good or doingthings to him he hates when he's bad. Dogs have reference levels for thingslike milk bones (high) or being jerked on a choke chain (low), and will do whatthey must to maintain those levels. There's nothing new about the way dogsreact to such things (part of the steaming pile of facts) - PCT models why theydo, and how. There are lots of other theories as to why - PCT is, so far, theonly plausible model of how - of the organization required.
More in the wild: there are a number of school teachers, principals,psychologists, and parents who are pretty happy with using PCT as an organizingprinciple to guide what they do and better understand the consequences, goodand bad, of one or another way of doing things. They feel it beats cookbookeducation courses a mile (if this happens, do that). One of the main thingsthey get out of it is the value of teaching children that they are controlsystems, and so are their classmates, and their teacher, and that everyone'sinteractions have to take that into account. It isn't a secret recipe forsuccess in the classroom; it's fundamental knowledge to be shared.
You have quibbled about us jumping from a few paltry controlledexperimental situations to conclusions about real life. You have now complainedthat we don't apply PCT to real life. Aside from that not being true, you can'thave it both ways. PCTers do experiments _and_ look at real life. The roots ofPCT lie in really close and detailed examination of real life. Any one of uscould show you four levels of your own hierarchical control organization(including a well-controlled reflex) in about two minutes.
You said: "I am focussing on design of control hierarchies to reflectactual behavior ... we should be able to define a HPCT system to cover anyknown behavior, whether or not the technology exists to actually implementit".
Yes. We should. We are in the process of doing that very thing. But youkeep bringing up examples that have already been taken care of, like reflexes.Why should we have to explain stuff on the net that's available in theliterature?
The journal that has "the blind men and the elephant" is Closed Loop, thejournal of the CSG. The same issue also has Bourbon's "Models and theirworlds". A bargain at $6. Send your check to me at 73 Ridge Place, Durango CO81301-8136.
Mary Powers
Date: Tue Jul 12, 1994 12:32 pm PST
Subject: Re: Replies to Paul and Jeff
From Tom Bourbon [940712.1341] >[Paul 940712 10:30] >>[Rick
(940711.2200)]
>> Could you give just one example of Skinnerian wheat?
> How about what you cited. Much of his work suggested that peoplecould be guided away from conflicting behavior (as externally perceived) bysetting up reinforcement structure that inclined their behavioral patternstowards 'better' patterns. Behavior (i.e. my perceptions of what you do) maybe modified by altering the inputs to you. This causes your perceptions to bedisturbed and perhaps adjust themselves so that my inputs resulting from youroutputs are closer to my reference values. There is of course a kind ofHisenberg effect involved. Skinner didn't put it this way, but the concept isthere.If you wish its 'truth' can be demonstrated by the fact that we cansuggest a control model for the observed results. In my experience childraising works, and I am certainly trying to modify my son's behavior. Trulytrying to control another in all things is very difficult and will causeconflicts, unless I can understand your control structure and control yourinputs (e.g. brainwashing or indoctrination).
But what you have described, Paul, is an example of one control system (A)disturbing a variable controlled by another (B) in such a way that A sees B'sactions matching a pattern that A wants to see. (I have numerous working modelsof pairs of PCT systems that interact in that way.) In that case, for A to"control" the actions of B, A must not prevent B from controlling the variablethat B intends to control -- in a very real sense, the actions of A are"controlled" by the fact that B is still in control of B's chosen variable. Bythe way, this situation sets up a beautiful opportunity for B to"counter-control" A's actions, while at the same time, A controls B's actions.It is indeed true that Skinner described similar instances of control andcounter-control, but he never understood why they worked as they did -- hethought it was a matter of stimuli controlling actions. He missed by amile.
Incidentally, you haven't really lived until you've exchanged control andcounter-control with a PCT model running as your partner, in "real human time"in a laptop computer. Come to one of our meetings and I'll give you anopportunity to do that. Or maybe I can send you a program and let you try it athome.
>> In principle, all controlled variables can be empirically shown toexist using The Test for the Controlled Variable.
> Eh, I'll hold off major comment until I can get hold of theliterature. However, it is not clear to me that because I can model the systemusing a control variable (or network thereof) that it follows that it is theonly mechanism that can work.
I think what Rick was talking about was what you had asked about -- theidea that there are empirical techniques for demonstrating whether _or not_control exists in a particular situation. If it does, there is no other modelthan a control theoretic model that can explain the observed phenomenon ofcontrol. (Notice I said control theoretic in the general sense -- PCT providesa control theoretic model that is set up in a particular way to emphasize theproperties of living systems, whereas most engineering versions of controltheory are set up with a different emphasis.)
> {note that I am impressed with the resolve shown to use only thissimple negative feedback mechanism to model all behavior; may yousucceed}
At least we will give it a good try.
> This appears similar to saying that since I can model control using adigital system that all life must be digital. Different functions can producethe same behavior. They are equivalent, not equal.
But that is not at all what we are saying -- we begin by saying that _if_control is found, _then_ only a control system can produce it. Can you suggestto us another way control might occur? No S-R system or plan-driven system canproduce reliable control in a variable environment, or can they work afterall?
>> What does "gestalt" have to do with anything?
> A set of control nodes containing controlled variables produce theobserved behavior, i.e it is the behavior of the system, not the parts. There_need_ not be a master variable. Distributed control systems do notnecessarily have to have 'master' nodes.
As Rick, what does this have to do with a "gestalt?"
>> If you agree that the behavior [moving the head in reaction to aswing] involves control then I suggest we put your claim to the test; let's seethe elaborations that you think are needed to make the model work.
> The example you use is familiar, being a martial artist. At a highlevel there is an observed 'desire' to avoid harm. This doubtless involvesavoiding pain. To avoid getting hit I can duck, sidestep, parry, or strikefirst interrupting the punch. Some network of control determines that a punchis coming, is likely to hit, and might cause harm. Some network of control'selects' the appropriate 'pattern' of behavior, probably by activating othercontrol loops 'programmed by long training. I am likely controlling many highlevel variables such as status (getting knocked down is losing face), form(pride in technique), principle of minimum force, etc. If I was justcontrolling for pain I would duck when anything hurt.
Paul, you are describing control as though it were a lineal processof:
input-->process ("determine")-->select appropriate preprogrammedaction-->act
I believe you left off that last step -- act. The system you describedwould not control anything; it would simply act in a pre-programmed way thatmight be suitable for the average value of similar input conditions it hadencountered in the past, but I'll guarantee you that such a system would neverwin the heavy-weight boxing title! I doubt that it's head would even be inplace after the first punch thrown by a challenger.
Also, we have been trying to tell you from the very start that thehierarchical PCT model includes many -- very many -- control loops runningsimultaneously in a system that is hierarchical and richly parallel. What'swith this idea you keep going back to -- the idea that there must be more thanone reference signal at a time. Of course that is the case, but sometimes onlyone reference signal is needed in a working model. When more are needed, theyare available.
Later, Tom
Date: Tue Jul 12, 1994 2:34 pm PST
Subject: Fact vs theory of control
[From Rick Marken (940712.1330)] Paul George (940712 10:30)
> Much of his [Skinner's] work suggested that people could be guidedaway from conflicting behavior (as externally perceived) by setting upreinforcement structure that inclined their behavioral patterns towards'better' patterns.
And this is just one of Skinner's observations that PCT shows to be acrock. 'Better' behavioral patterns means behaviors that are closer to thereference specifications of the observer (like Skinner). The behavior patternsthat meet the reference specifications of an observer are unlikely to produceperceptual results that are "better' for the organism itself.
> Behavior (i.e. my perceptions of what you do) may be modified byaltering the inputs to you. This causes your perceptions to be disturbed andperhaps adjust themselves so that my inputs resulting from your outputs arecloser to my reference values.
That's one very hopeful "perhaps".
Me:
>> In principle, all controlled variables can be empirically shown toexist using The Test for the Controlled Variable.
You:
> it is not clear to me that because I can model the system using acontrol variable (or network thereof) that it follows that it is the onlymechanism that can work.
This is not the point of what I said . You are having a problem, I think,distinguishing the FACT of control from any THEORY that might be proposed toexplain it. The Test for controlled variables establishes the FACT that avariable is under control. In our tracking experiments, for example, the factthat a person's actions precisely counteract disturbances to the position ofthe cursor demonstrates the FACT that the position of the cursor is acontrolled variable. The Test can be used to establish that any other variableis under control as well; reinforcement rate, for example. The existence ofcontrolled variables is a FACT; there is no theory involved.
We are not saying that PCT is a good theory because we can "model thesystem using a control variable". We are saying -- HEY LOOK, THERE ARECONTROLLED VARIABLES ALL OVER THE PLACE AND LIVING SYSTEMS ARE RESPONSIBLE FORTHEIR EXISTENCE. This is a fact to be explained, just as it is a fact thatobjects accelerate to earth at 32ft/sec^2. The existence of controlledvariables is a fact that was not observed until Powers pointed it out -- justas linear acceleration is a fact that was not observed until Galileo pointed itout.
Newton provided the theory that explained Galileo's (and Kepler's)findings; the theory that explains Powers' findings was already in existence(and Powers knew it); control theory.
> {note that I am impressed with the resolve shown to use only thissimple negative feedback mechanism to model all behavior;
So now you see that this is not quite a correct statement; we use negativefeedback control to model the FACT of control behavior. "Simple" feedbackcontrol is the only model we know of that explains the fact of controlledvariables. If you (or anyone) knows of another model that explains the samefact (control) then, by all means, show us.
PCT doesn't have a problem of acceptance because people don't like theTHEORY; PCT has a problem of acceptance because people don't know what FACTHPCT has been developed to explain -- the FACT of control.
Moreover, psychologists have no idea that CAUSING something and CONTROLLINGit are two completely different things. I don't think there are manypsychologists, for example, who would recognize that a statement like"reinforcement controls behavior" is demonstrably ridiculous. Reinforcementswon't act to resist disturbances that move behavior from its referencestate.
Me:
>> If you agree that the behavior [moving the head in reaction to aswing] involves control then I suggest we put your claim to the test; let's seethe elaborations that you think are needed to make the model work.
You:
> I hope you see that what I mean by elaboration of the model is anelaboration of the set of controlled variables , reference values, and thecontrol network; not the model of perceptual control.
Then I agree that these elaborations are necessary; they are already partof the HPCT model. I know that control of almost any variable really involvescontrol of a hierarchy of controlled variables; control of reinforcement rateinvolves control of (perceptions of) muscle tensions, limb configurations,transitions between configurations, sequences of transitions, and many othervariables that result in a particular rate of reinforcement. But, forsimplicity, we can often lump all these variables together to produce a simplemodel of control of one variable -- rate of reinforcement. We do not imaginethat a rat, for example, is just a furry thermostat with one sensor and oneoutput device. If you have this impression of PCT, I think it would be quicklydispelled when you _read_ "Behavior: The control of perception". HPCT is onevery rich model; "simple feedback system" doesn't really capture it.
> At worst I would change "a perceptual variable" to "a set ofperceptual variables".
It's a deal! Even the simplest organisms control thousands (really) ofperceptual variables simultaneously; and many of these perceptual variables arecontrolled as the means of controlling other perceptual variables. That'shierarchical perceptual control theory. It looks kinda like an Albus model onpaper; the main differences are 1) HPCT is explicitly a model of control (thefact of the existence of controlled variables) and 2) the variables that arecontrolled in PCT are PERCEPTUAL variables.
Best Rick
Date: Tue Jul 12, 1994 4:04 pm PST
Subject: Re: Replies to Paul and Jeff
[Paul George 940712 15:40] >Tom Bourbon [940712.1341]
> But what you have described, Paul, is an example of one controlsystem (A) disturbing a variable controlled by another (B) in such a way thatA sees B's actions matching a pattern that A wants to see....
Deliberately. I attempted to describe in PCT terms (or a reasonablefacsimile ;-) a useful idea from Skinner. Hoped I might demonstrate that I havesome inkling of what you are talking about.
> It is indeed true that Skinner described similar instances of controland counter-control, but he never understood why they worked as they did -- hethought it was a matter of stimuli controlling actions. He missed by amile.
That he didn't 'properly understand' the mechanism is to me of littleimportance - that is the chaff. I'm just not a purist or true believer bynature. And as I commented to Bill P. today, I think he meant that all youcould observe was stimuli apparently producing actions, and that was sufficientfor behavioral modification.
>> A set of control nodes containing controlled variables produce theobserved behavior, i.e it is the behavior of the system, not the parts. There_need_ not be a master variable. Distributed control systems do not necessarilyhave to have 'master' nodes.
> As Rick, what does this have to do with a "gestalt?"
We seem to have very different definitions for the term. I mean somethingthat does not appear until a 'critical mass' of components exist. Systems showgroup behavior, and it is sometimes a step function of complexity. (N-1)components won't do it, and there must be the right N components. The littleman doesn't work until all three control functions exist, the sensory inputsexist for feedback, the output mechanisms exist, and all are properlyinterconnected.
> Paul, you are describing control as though it were a lineal processof:
> input--process ("determine")--select appropriate preprogrammedaction--act
> I believe you left off that last step -- act.
Nope, while action and input may be lineal, control is usually continuousor at least periodic. (I vaguely seem to recall some medical research thatindicated that sensory inputs such as vision had some kind of sense>processcycle in the brain with a detectable period). I considered the 'selection' thesame (from the control system's point of view) as acting. The act is usuallydoing something to allow another series of nodes to 'take action'.
> What's with this idea you keep going back to -- the idea that theremust be more than one reference signal at a time. Of course that is the case,but sometimes only one reference signal is needed in a working model. Whenmore are needed, they are available.
It had been stated as the general model, but every time an example wasgiven it was in terms of 'a controlled variable' as if each loop had only one,usually refering to something complex. When I suggested a more complex set, Iwas told it was not needed. This produced confusion. I am relieved that we weresaying the same thing (?) and that my initial understanding was correct.
Hope all these replies aren't eating up too much bandwidth.
Paul
Date: Tue Jul 12, 1994 4:07 pm PST
Subject: Re: Fact vs theory of control
[Paul George 940712 15:50] >[Rick Marken (940712.1330)]
Most of my comment is in my previous post to Tom on the samesubject.
> 'Better' behavioral patterns means behaviors that are closer to thereference specifications of the observer (like Skinner). The behaviorpatterns that meet the reference specifications of an observer are unlikely toproduce perceptual results that are "better' for the organism itself.
If that were purely true neither society nor parenting would be possible.Since a fair amount of our perceptual variables are hardwired, there arecommonalities, and there can be complementary goals.
> This is not the point of what I said . You are having a problem, Ithink, distinguishing the FACT of control from any THEORY that might beproposed to explain it. The Test for controlled variables establishes the FACTthat a variable is under control..... We are saying -- HEY LOOK, THERE ARECONTROLLED VARIABLES ALL OVER THE PLACE AND LIVING SYSTEMS ARE RESPONSIBLE FORTHEIR EXISTENCE. This is a fact to be explained....
I don't think I dispute the fact. Sometimes I question the theory or itsmodeling. The capitalized point is the basis of my interest in PCT, I tooobserved it a long time ago (but I don't publish, no credentials ;-)), though Ithink in terms of perceptual models (i.e a control network or set ofvariables)) The only question I have is whether postulating and modeling aparticular variable(or set thereof) establishes _in all cases_ that it is infact the correct one used by an organism apparently demonstrating the modeledbehavior. You (Rick) seem to be a little more dogmatic on the question thanmost others. I have interpreted most of your examples of "the test" as'proving' a particular control variable, not the existence of control. In yourexample the cursor position could be the variable, or it could affect thevariable. The cursor could be perceived in a number of ways, and 'position' isnot a simple concept. At any rate this is a bit of a fine distinction, and addslittle insight. It is fine to look at position as if it were the variable. Itworks to describe the control behavior.
Paul
Date: Tue Jul 12, 1994 4:33 pm PST
Subject: Powers Responses.
[Paul George 940712 17:00] >[Bill Powers (940712.0745 MDT)]
> Skinner would vehemently have denied that he had goals "in mind", ofcourse, saying "the environment made me say that."
Umm, It's your field not mine, and it has been a long time since I've readhim. I interpreted him to be saying that we couldn't determine what washappening within the organism, and so should focus on what was without as wecould observe it. And in a sense the environment does 'make us do it' as thatis what we perceive and indirectly attempt to control. Proponents of a schoolsometimes make extreme pronouncements to 'stake out the territory'. Their otherstatements and actions sometimes reveal they really don't believe them.
> The problem is that this "decomposition" assumes that you alreadyknow what the parts will be. There is nothing in the Big Picture of washingthe car that tells you what you need to do in order to achieve thatresult....The way you decompose the big picture depends on what you alreadyperceive about the actual situation at the time the main task is to be carriedout.
When decomposing behavior perhaps. When we do it we are trying to determinewhat we are talking about(general to specific) and what the components need tobe. It is a 'divide and conquer' strategy. When we get to a set of concretecomponents or features we usually need to totally reorganize thehierarchy.
OTOH when I do process (a.k.a. activity) modeling, tasks have preconditionsthat are evaluated to determine if the task needs to be performed.Uponcompletion postconditions to affect the activation and processing of othertasks. The model tries to capture all potential tasks. Executing the modelinvolves traversing appropriate paths. I have no idea how this might apply toPCT, though it could be a mechanism for dynamically reorganizing or activatingcomponents (heresy I know) of a control network.The evaluation of conditionscould be viewed as an intermittent or reactive control loop.
> Ah, then I take it you have obtained Rick Marken's book, "Mindreadings," and have seen his article called "The blind men and theelephant,"
Nope, the parable predates it somewhat :-), and applies to a lot of what Ido. I do however look forward to reading it.
>[from Mary Powers 940712]
> It's a little difficult to know where you stand ;-)
In answer, at this point I don't stand I just raise questions and examplesfrom my experience. I consider most of my points or observations to be minor,usually on how things have been phrased, or the exact contents of a givencontrol model. Occasionally I question the logical necessity of a conclusion orextrapolation. I try to feed back my perceptions of what was said so I canevaluate my interpretation. When I get hold of the published literature andhave the oportunity to study it in detail, then I'll take a stand. Y'allproject a lot more certanty upon me than I have, and assume a lot strongercriticism than is there (apparently because of historical attacks with similarpatterns). I use words like 'may', 'appear', 'seem', and 'not sure' quitecarefully. As the old saw goes "I know you think you understand what youthought I said, but you may not realize that what you heard was not what Imeant". Communication is a tricky thing, particularly without consistent setsof semantics.
> You have quibbled about us jumping from a few paltry controlledexperimental situations to conclusions about real life. You have nowcomplained that we don't apply PCT to real life. Aside from that not beingtrue, you can't have it both ways.
I said that they were minor factors, at worst logical leaps. "Not Proven"is not the same as rejected. I have acknowledeged repeatedly that PCT is andhas been applied to real life, and derived from observations about it. In myview that is the beauty of PCT, that it has both high level application and lowlevel (demonstrable) mechanisms. My 'quibble' is a repetition of the frequentstatements, including yours, that both ends haven't _yet_ met in the middle interms of executable models. That simply is what I consider the set ofinteresting problems to be addressed.
> But you keep bringing up examples that have already been taken careof, like reflexes.
Based upon what I have seen (my fault not yours) you have not 'taken careof it', you have addressed it. You have built models that demonstrate that itis possible for control systems to demonstrate equivalent behavior to anorganism. A true accomplishment. But it still doesn't follow that it is in factthe only possible model, or the mechanism in fact used by living systems at alllevels. The fact that the other current behavioral models are even lessvalidated (or not at all) doesn't to my mind change things. I may accept yourhypothesis or conclusions without agreeing that they are proven orinescapable.
> Why should we have to explain stuff on the net that's available inthe literature?
Because the literature is not readily available in public libraries orbookstores. As you are an unpopular school of thought, they are not evenpresent in many university libraries (which in any event not everyone hasaccess to). You don't (according to the FAQ) provide Closed Loop or otherpapers at the ftp site. Until I can arrange to borrow or afford to buy thematerial, I must rely on what Bill Powers, Marken, and Tom Bourbon say in theirposts. And BTW they are very clear writers and likely successful educators. Ihave learned a lot about their thinking and technical aspects of PCT. Theirposts are informative, if not always directed towards what I thought I wasbringing up.
I am trying to get BCP via inter-library loan, and will order CL from youshortly (though the issue number might be helpful).
Thanks, Paul George
Date: Tue Jul 12, 1994 4:59 pm PST
Subject: Re: Testing models
[Paul George 940712 15:00] >[John Anderson (940712.1330)]
> Paul, this seems contradictory. First, you appear to say that youagree with Bill Leach's statement that perceptions do not have to be undercontrol, and then you turn around and say "We perceive nothing that we are notcontrolling for". Or do I misunderstand?
May be a terminology problem. I think I agree with Bill. If I understandthe PCT basics correctly, we do not perceive anything but controlled variables(though I suppose a noop control loop is conceivable). The nature of ourperceptions change only as a function of changing variables or associatedreference values (or possibly comparison algorithms). At a given point in timea given variable may be set within reference bounds or no sensory inputs havecaused it to perturb. At such times no actions are being generated and so thecontrol loop could be considered as not actively controlling the perception. {Ithink this matches [From Bill Powers (940712.0745 MDT)]} I can't say if Bill Lwould consider the perception as being controlled or not in this situation. Ifwe allow a control loop to hibernate or in some way prevent error signals frombeing acted upon, then too we are not controlling the perception.
Another interpretation is possible is if you are allowed to distinguish'perception' variables from 'controlled' variables. Either perceptual variablesare processed in some way to 'compute' the values of controlled variables (e,g,a trend, upper/lower limit, etc.), or the latter are a subset of the former. Inthe first case the perceptions are controlled indirectly, and in the second notall perceptions are controlled at a given time. Can't say how formal PCT judgesthe situation. I'm not sure either elaboration is needed.
Paul
Date: Tue Jul 12, 1994 7:42 pm PST
Subject: Say what???
[From Rick Marken (940712.2030)] Paul George (940712 15:50)
> The only question I have is whether postulating and modeling aparticular variable (or set thereof) establishes _in all cases_ that it is infact the correct one used by an organism apparently demonstrating the modeledbehavior.
What do you mean by "in all cases"? I have no idea what you could possiblymean. We establish by Test that an organism is controlling a particularvariable in a particular situation; if The Test reveals that reinforcement rateis a controlled variable then it is -- IN THAT CASE. The next step is to builda model that can also control that variable. What in the world is your conceptof a controlled variable??
> You (Rick) seem to be a little more dogmatic on the question thanmost others.
On what question?? Actually I think Bill and Tom can match me dogma fordogma. I'm just a little stricter (high gain control) than they are -- thoughTom is making great strides ;-)
> I have interpreted most of your examples of "the test" as 'proving' aparticular control variable, not the existence of control.
I would feel a lot more comfortable about your interpretations of myexamples if you would call a controlled variable a "controlled variable" ratherthan a "control variable". The term "control variable" implies that a variableis doing the controlling. In control, one of several different variables, (byvirtue of its position in a negative feedback loop) is _controlled_; there isno single variable that could be called the "control variable".
And how in the world could one demonstrate the existence of a controllEDvariable and not be demonstrating the existence of control?
> In your example the cursor position could be the variable, or itcould affect the variable. The cursor could be perceived in a number of ways,and 'position' is not a simple concept.
Position may not be a simple concept in some way or other, but its prettyeasy to measure. In compensatory tracking tasks it's just the horizontalposition of the cursor on the screen, in pixels. Call this variable x. In ourexperiments x = h + d where all variables vary over time. Position, x, is atany instant the sum of a disturbance variable, d and the position of a"control" handle, h. If, over 2000 or so samples of each variable, the varianceof x is virtually zero because h = -d (approximately) then the x variable is acontrolled variable. It's true that the actual controlled variable (theperceptual representation of x) could be any monotonic function of x, but thatwould have to be determined by other test. There is no question that x is aVERY GOOD approximation to the actual controlled variable. The nearly perfectrelationship between h and d is proof that x (or some monotonic functionthereof) is under control -- it is a controlled variable.
Am I missing something?
Best Rick
Date: Tue Jul 12, 1994 7:53 pm PST
Subject: Re: Fact vs theory of control
[Avery Andrews 940713.1335] (Paul George 940712 15:50)
> The only question I have is whether postulating and modeling aparticular variable(or set thereof) establishes _in all cases_ that it is infact the correct one used by an organism apparently demonstrating the modeledbehavior.
I'd say that it doesn't, but that it leaves you with a hypothesis that canbe further challenged and investigated. In this respect I don't think PCT isdifferent from any other empirical field of inquiry. We can be wrong about whatvariables an organism is controlling, and also about how apparent control isachieved. For example, perhaps one of the `transactionist AI/Alife' people(Beer, Brooks, Horswill, etc.) will come up with an architecture that produceswhat appears to be control by some means completely different from thatproposed in PCT. Then we can look for evidence as to whether living systemsemploy this architecture or not.
Fowler and Turvey, somewhere (I think in a book edited by Stelmach in thelate 70's) noted that respiration rate looked like it might be under feedbackcontrol, but argued that it wasn't. But an important point of their argument isthat it doesn't really act like a control system, but like a dynamic systemwith a whole series of attractor basins, so that a sufficiently largedisturbance will push it from one to another.
It might in fact be difficult to distinguish a multilevel control systemfrom some random kind of dynamic system - since lower level reference signalsare derived from higher-level error signals (plus perhaps perceptions), youmight get attractor-like behavior as a system shuffled around between differentlower level goals in an attempt to satisfy a higher level ones. But then PCTpredicts that there will be a higher-level goal that you might succeed inidentifying, whereas dynamic systems theory says nothing, and so is, I think,inherently dead-ended.
Another problem in evaluating human performance is that there can beseveral control systems operating simultaneously, at different time-scales.E.g. I can be controlling for getting my sailboat from Manchester Harbor toSinging Beach, but doing it at the moment by controlling for heading towardssome island, but then I discover that it's the wrong island, so this referencelevel is replaced by some other one. Sorting out what's really going on isobviously not going to be easy, but I think that this is a difficulty inherentin the subject matter (living systems), rather than a nasty feature of theapproach (PCT).
Avery
Date: Tue Jul 12, 1994 10:07 pm PST
Subject: Re: Fact vs theory of control
<[Bill Leach 940712.20:44 EST(EDT)] >[Paul George 94071215:50]
Paul, I'm not sure if you picked up on the most significant thing that Rickwas saying there:
Negative feedback control theory is the ONLY known explaination forbehavior. That is the point and that is the reason for PCT.
You mentioned yourself (several times I believe) that "control theory couldexplain a certain behavior" and added that "you were not convinced that controltheory could be the only explaination".
Turn that around a bit for a minute...
S-R (stimulus-response) actually does appear to explain some behavior (atleast under some conditions).
PCT predicts that S-R will appear to work under the "right" conditions butwhere the two differ is that PCT works to explain the "anomalies" that appearin S-R testing instead for just dismissing them.
An additional thought is that some behavior quite clearly can only beexplained by control theory unless one is willing to postulate a "theory" thatrequires different explanations for every change in behavior.
When one recognizes that living systems do indeed exhibit control systemoperation in testable situations and that control theory is the only theorythat can explain the manner in which behavior is actually seen to vary asenvironmental conditions change AND that there is an ever mounting body ofphysical evidence to support that there are actual physical control structurespresent in living beings then it seems reasonable to me to contend that untilproven otherwise, PCT is the proper approach to understanding.
Much of the difficulty that you are presently experiencing on the net will"go away" when you have had a chance to read some of the literature. This isespecially true because you will see how most of what we tend to believe as"characteristics" of humans is rather learned traits.
Also, you will see where many things that humans do that are thought to beextremely complex could in fact be quite simple if performed by a controlsystem with certain references set.
The literature will show that a great deal of consideration has gone intounderstanding "general human nature" already and that many of the conclusionsare surprising.
-bill
Date: Tue Jul 12, 1994 10:16 pm PST
Subject: Re: top-down, bottom-up, etc.
<[Bill Leach 940712.20:24 EST] >[Bill Powers (940712.0745MDT)]
Not to belabor a point but...
Is the statement; "Behavior results from the control of perception."imprecise? Is there really a reason for not stating it in such a manner?
Or for that matter, again what specifically is wrong with saying;"Behaviors is the result of controlling perceptions."?
I am not trying to be contrary here but rather I really do want to know thereason the above is not a desirable way of stating the case.
-bill
Date: Wed Jul 13, 1994 7:27 am PST
Subject: Re: Fact vs theory of control
[Dan Miller (940713)]
To Rick Marken:
Hi, again, Rick. In your tutorial with Paul George you made the followingassertion:
> The existence of controlled variables is a FACT; there is no theoryinvolved.
There you go again with FACTS. Do you mean to say that facts have anexistence without theory, hypotheses, or epistemology? If this is true, thenall we scientists have to do is reveal the facts of the universe.
I do not disagree with your perceptions of FACTS, but didn't those beforeGalileo and Kepler think that the Earth was the center of the universe. Weren'ttheir observations FACTS? Now we understand that their model was invalid (itdidn't fit all the observations). Of course their fate will not be ours. Iwould hesitate to say that we have "immaculate perceptions." However, anelement of doubt might not hurt.
As for FACTS. If facts are perceptions (shared and reproducible), then theymust be attached to theory (broadly considered). As conscious perceptions wesee them (feel, etc.) and think of them in certain ways - temporally and oftenspatially separated from the situated perceptions themselves.
Later, Dan Miller
Date: Wed Jul 13, 1994 7:55 am PST
Subject: Re: Say what???
Paul George (940713 0930)
I think we have a common understanding of control and testing forit.
>[Rick Marken (940712.2030)]
> What do you mean by "in all cases"?
Always x itself. The variable I test must be the actual variable. Sometimesit is x, and sometimes a function of x, as you say below. Sometimes (though notin your examples) x could be a function of the true variable(s). We can but tryto find a 'metric' that lets us determine something useful about the perceptualfunction.
> And how in the world could one demonstrate the existence of acontrollED variable and not be demonstrating the existence of control?
You couldn't, and I didn't intend to ever suggest that you could or wouldwant to.
> I would feel a lot more comfortable about your interpretations of myexamples if you would call a controlled variable a "controlled variable"rather than a "control variable".
No problem, though that is the term process control system designers use.Ack on the semantic distinction. Much of the interaction of the last couple ofweeks is reaching a common set of terminology.
> Am I missing something?
Don't think so. Don't think I am either, at least on your particular areaof focus. I'll grant that 'strict' or 'rigorous' is a better characterizationthan dogmatic.
Please note that I don't usually comment on things I agree with inanother's post in the interest of saving bandwidth. Silence is acknowledgementor agreement. My view of a forum is constructing an artifact in cyberspace isthe product of all posts.
Later, Paul George.
Date: Wed Jul 13, 1994 7:55 am PST
Subject: Artificial
Paul George 940713 10:00 >[Avery Andrews 940713.1335]
> For example, perhaps one of the `transactionist AI/Alife' people(Beer, Brooks, Horswill, etc.) will come up with an architecture that produceswhat appears to be control by some means completely different from thatproposed in PCT.
Has anyone in the CSG taken a close look recently on the work on ArtificialLife, Core Wars, or Genetic Algorithm software?
I haven't looked at it in detail for a while, and never from the standpointof formal control theory. However, these things do seem to demonstrate behaviorsimilar to ecologies. Simple strategy has the effect of goal seeking, withoutit in fact being programmed. The observer or programmer intends to maximizesurvival, but the program does not explicitly have such a goal. Intent andeffect need not be the same thing, and a successful survival strategy will tendto persist. This corresponds to an attempt to model in software ethology fromthe behavioral science domain.
Comments?
{This is not criticism, nor challenge, just a question}
Paul George
Date: Wed Jul 13, 1994 9:07 am PST
Subject: Paul George 940713 10:00
Thomas Baines ???
Reference the use of genetic algorithms in control, look at the stiles andGlickson article in IBM JOURNAL OF RESEARCH & DEVELOPMENT, Vol. 38, No. 2,p. 157.
Although they are dealing with a "Highly parallelizable rout planner basedon cellular automata algorithms", the translation to search and control is nota biggie.
Date: Wed Jul 13, 1994 9:20 am PST
Subject: Controlling Skinner
[From Rick Marken (940713.0900)]
I said:
> 'Better' behavioral patterns means behaviors that are closer to thereference specifications of the observer (like Skinner). The behaviorpatterns that meet the reference specifications of an observer are unlikely toproduce perceptual results that are "better' for the organism itself.
Paul George (940712 15:50) --
> If that were purely true neither society nor parenting would bepossible.
It is purely true, and society and parenting are still possible (andactual) because neither is based on the control of other people's behavior.When people seriously apply Skinnerian methods to building societies or toparenting children the result is always disaster. People have always tried touse Skinnerian methods -- even before Skinner -- but they can't apply them forlong because they almost always place the "applier" of these methods in aconflict - - with him or herself and/or with the controllee. My wife and Isuccessfully (to our minds) parented two kids without ever resorting toSkinnerian nonsense (the down side is that I now I enjoy eating all thoseM&Ms that we didn't need to use as reinforcers;-))
PCT shows that it was natural for Skinner to have wanted to be in control;all people are controllers; controlling is a good thing. The problem is thatyou run into trouble when you try to control other controllers -- especiallyother controllers who can perceive and control the same variables that you can.Skinner just never figured out that other people are just like him -- they arecontrollers. Skinner was so wrapped up in his controlling that he neverrealized that what he was doing (controlling) was the interesting phenomenon;not the results of his controlling (like pigeons playing ping pong).
Best Rick
Date: Wed Jul 13, 1994 11:47 am PST
Subject: Facts, Theories & Illusions
[From Rick Marken (940713.1030)] Dan Miller (940713)
> There you go again with FACTS. Do you mean to say that facts have anexistence without theory, hypotheses, or epistemology?
I think I'll have to say "yes".
> If this is true, then all we scientists have to do is reveal thefacts of the universe.
I don't see why that's true. I think that one aspect of science is"revealing" or describing facts; another is trying to make sense of them -- tofigure out why these facts exist. That's where theory comes in. Ithink.
> I do not disagree with your perceptions of FACTS, but didn't thosebefore Galileo and Kepler think that the Earth was the center of the universe.Weren't their observations FACTS?
I don't think that they "observed" the earth to be in the center of theuniverse; they observed (as you still can) that the earth is stationary, thatthe sun, moon, stars and planets move across the sky at different rates, and soon. A universe with the earth at the center was their explanation (theory) ofwhy they saw what they saw; it was an explanation of the facts. Theirobservations were (and still are) facts.
I don't think theory determines facts; it determines how we interpret thefacts. For example, I personally believe that the sun is the center of thesolar system, and that the earth rotates on an axis as it moves around the sun.But I still observe the same facts that the ancients observed; I perceive theearth as stationary with the sun, moon, stars and planets moving across the skyat different rates. I just think that the reason why I see these facts isdifferent than the reason assumed by the ancients.
> Of course their fate will not be ours. I would hesitate to say thatwe have "immaculate perceptions." However, an element of doubt might nothurt.
Our model will ultimately be replaced -- as was theirs. But the facts willremain (as did theirs). Our perceptions (aided or unaided) are the same asthose of the ancients -- it's our explanation of the cause of those perceptions-- our theory -- that will get replaced.
> As for FACTS. If facts are perceptions (shared and reproducible),then they must be attached to theory (broadly considered).
I guess I disagree with this. If facts are perceptions, then I don't seewhy they MUST be "attached" to a theory. In fact, the idea of theorizing toexplain facts is rather new -- starting in the 1500s or so. Perceptions arejust perceptions -- some can be controlled; some can't. It's a relatively weirdbreed of cat (a scientist cat) who finds it necessary to build models(theories) to explain the reality that is presumed to lie beyond what weperceive. Western science proves that such models can provide a powerful toolfor controlling perceptions; but, as Paul George pointed out, we are able tocontrol most perceptual variables without having any understanding of how weare able to do it.
By the way, PCT does not deny that certain facts exist; the statisticalresults observed in the behavioral sciences still count as observations (thoughthey are often so unreliable that we don't like to call them "facts"). Still,the fact that we do observe statistical relationships (across subjects) betweenvariables like the color of the room and the productivity of the workers in theroom (to pick a weird "fact" off the top of my head) is not denied by PCT; PCTjust says that the REASON for this fact (observation) differs from what hasseemed like the obvious reason -- that there is a causal relationship betweenwall color and productivity.
I think theory influences how we interpret the facts; it also influenceswhat facts we consider important and what facts we attend to; but I don't thinktheory influences what the facts are. At least, I don't think it should.
I agree that I am not always clear in how I deal with the relationshipbetween fact and theory. For example, I often refer to the S-R relationshipsobserved in the behavior of living organisms as an "illusion". This suggeststhat these relationships (in light of the PCT model) don't exist; that thetheory (PCT) revokes what was thought of as a fact (S-R relationships). But theS-R relationships are facts; PCT doesn't change that. The "illusion" is not theS-R relationship itself -- that is a fact. The illusion is that there is acausal relationship between S and R mediated by the organism; the illusion isnot that S-R is a fact; the illusion is the theory typically assumed to explainthat fact. Saying that S-R is an illusion is like saying that the sun movingacross the sky is an illusion. The perception of the sun moving across the skyis NOT an illusion; it is a fact. The illusion is the theory -- that it is thesun that is moving. The right theory (currently) is that the earth is rotatingand that the sun just appears to move as we rotate past it. So maybe the word"illusion" is inappropriate when referring to incorrect theories of the facts.Mea culpa.
Best Rick
Date: Wed Jul 13, 1994 1:11 pm PST
Subject: 6=+3g~},w_/_/{w{{^{o
[From Bill Powers (940713.1230 MDT)] Paul George (940713)
Paul, I think some of our semantic difficulties need clearing up by way ofa diagram. Below is our standard diagram of a control system with verbaldefinitions of the components and signals. Would you attach to this diagram thelabels you are accustomed to using for the parts of a control system? It seemsto me that a process-control engineer uses the term "control variable" to meana variable that causes control actions, while we use "controlled variable" tomean a variable that is altered by actions and, via the closed loop, iscontrolled by them. There may be other problems; for example it is customary incontrol engineering to speak of the reference signal (in our diagrams) as the"input."
I will present this diagram on its side (as compared to the way we usuallydraw it), so it will look more like standard engineering diagrams. Higherlevels would then be to the left, and systems operating in parallel at the samelevel would be stacked vertically. You might want to insert blank lines here sothe following will appear on one page.
:
:
B ======
=========>|| F2 ||---C----------D
|| ====== |
^^ : v
|| : =====
====== : || F3 ||
A ===>|| F1 || : ======
====== : | E
^^ : | |
|| H ====== v ====== |
====<======|| F4 ||<-----G------ F <----|| F5||<---
====== ======
:
:
The vertical dotted line separates the control system (to the left) fromthe physical (observable) environment outside it (to the right).
Let's start with the environment part, to the right of the vertical dottedline.
C = path by which action of the system (the output of F2) affects
D = an observable physical variable. F2 is an output transducer.
The state of D has an influence on
F = another physical variable, via
F3 = a function representing physical laws connecting D to F.
There is in general a physical variable
E = variable that independently contributes to the state of F via
F5 = a function representing physical laws connecting E to F. (There may bemany independent variables like E, contributing to the state of F via differentfunctions like F5).
Note that F = F3(D) + F5(E). Also, C and G have no special names in PCT;they are pathways.
Now we come to the control system.
At the input interface, there is a function F4 which is affected via
G = a pathway through which the variable F physically affects the system'ssensors, and produces an output
H = a signal standing for the state of F.
F4 can contain not only sensors, but computations on the values of outputsfrom sensors.
The signal H enters the function
F1 = a function which subtracts H from
A = a signal that has a state equal to the desired state of H, andproduces an output
B = a signal that is a measure of the discrepancy between A and H.
The signal B enters
F2 = an output transducer which converts the signal B into
D = the physical variable already mentioned, via
C = a pathway connecting the output of F2 to the variable D.
These are the functions, signals and variables that comprise the basiccontrol system as seen in PCT. Engineering diagrams generally leave out onemore of these elements, or lump several together, but they are always there inreal closed-loop control systems.
A comment on the environment part of the model is needed. In fact, thevariable D might be multiple physical variables, affected by many pathways C;there might be many functions F3 connecting the effects of the system to manyvariables F affected by many independent variables E. The variables F mightaffect many sensors through different paths at the input to F4.
The output of F4, however, is always a scalar variable, the value of thefunction F4(F[i]) for all i. In a real system there may be many functions F4,each producing a signal H that is a different function of the same set ofsensor inputs. Each F4 would be part of a different control loop acting inparallel and sensing a different aspect of the collection of physical variablesF. Each such loop would be drawn as a separate control system and placedvertically above or below the system shown above. A given F4 determines whatfunction of the detailed set of environmental variables F is represented byH.
-----------------------------------
OK, now what we need from you is a set of labels to go with A..H, andF1..F5. Each signal (inside the system), variable (outside the system) andfunction has a specific name in PCT. These names are always used in the sameway to mean the same part of a control system or its environment. If we can setup a translation table, it should be possible to translate a PCT description ofcontrol into a description in terms of any other engineering subculture. Howabout giving it a try?
Best, Bill P.
Date: Wed Jul 13, 1994 1:35 pm PST
Subject: Facts
[From Bill Powers (940713.1300 MDT)] Dan Miller (940713) --
Butting in:
> There you go again with FACTS. Do you mean to say that facts have anexistence without theory, hypotheses, or epistemology? If this is true, thenall we scientists have to do is reveal the facts of the universe.
The answer isn't a simple yes or no. There's a progression from simplyrecording the occurrence of an experience (brightness exists now) to readingall kinds of inferences into such experiences (somebody is shining a flashlightinto my window).
In the old two-psychoanalyst joke (How are you --- what did he mean bythat?), the joke has to do with an unwarranted model-based perception ofimplications that take a theory even to detect. But if the second psychoanalystsaid "I heard him say 'How are you'" there would be no joke, and indeed nopoint. It is a fact, as we think of facts, that the first psychoanalyst said"How are you?". The inter-rater reliability of observers asked whether thissentence was spoken with an interrogative inflection (one of them) would beextremely high. Even a person who spoke only a different language might be ableto identify that sentence among tape-recordings of different sentences anddeclare that recording 17 is the one that was spoken.
The relationships we observe in PCT which identify the existence of controlare, as nearly as possible, incontrovertible observations. We don't often gothrough all the details because we're so familiar with them, but we could. Iapply a force to an arm. I ask the person who owns the arm, "Am I applying aforce to your arm?" The answer would be "Yes." I say "I feel resistance to mypush; do you feel yourself resisting it?" The answer would be "yes." I theninject curare into the person's muscles, and try again. This time there is noresistance to my push. So I say that what I am observing meets the criteria forthe phenomenon I call control, or at least some of the main ones.
The Test for the controlled variable, while it is used in PCT, is nottheory-based in the sense of depending on control theory to work. The theoriesit depends upon are those of physics, and they are among the oldest andbest-supported theories known to human beings. A force applied to an objectwill cause that object to accelerate. If one force is applied, but the objectdoes not accelerate according to A = F/M, then there must be a second forcebeing applied in the opposite direction so that the net force is zero. When wefind the second force, we observe that it is equal in magnitude and opposite indirection to the first force. So we accept this aspect of control as havingbeen demonstrated as a fact.
There is one kind of fact that is, as far as I can see, absolutelyreliable. If I see a purple dog floating three feet off the floor, all kinds ofdisputable statements can be made about this experience, such as "there isn't(or is) really a dog there" or "the dog that is floating there isn't reallypurple." But there is one indisputable statement that I and I alone can make: Isee a purple dog floating three feet off the floor.
That to me is a ground-zero bedrock Fact. It is a description of experiencedevoid of opinion about it. The moment you get away from that level of Fact --the moment you say anything ABOUT that fact other than simply reporting itspresence -- interpretation, opinion, and theory get into the act.
The aim of experimental science is to provide facts that are as near aspossible to bare reports of experience. A true experimental scientist wouldnever report that an animal presses a bar at x presses per minute and gets yreinforcements per minute. It does not get reinforcements; it gets little brownfood pellets. This scientist would never say that 10 reinforcements per minuteis enough to maintain 150 bar-presses per minute. The report would merely saythat when the rate of delivery of little brown food pellets was 10 per minute,the rate of bar-pressing was 150 per minute. "Maintaining" is not a fact, butan interpretation.
Best, Bill P.
Date: Wed Jul 13, 1994 2:02 pm PST
Subject: Behavior and control
From Tom Bourbon [940713.1512]
>[Bill Leach 940712.20:24 EST] >>[Bill Powers (940712.0745MDT)]
> Not to belabor a point but...
> Is the statement; "Behavior results from the control of perception."imprecise? Is there really a reason for not stating it in such amanner?
> Or for that matter, again what specifically is wrong with saying;"Behaviors is the result of controlling perceptions."?
Bill L., I think what Bill P. is getting at is that behavior is notsomething that (merely or just) "results from" the control of perception. Ifthat were the case, then we would need to identify what it is, other thanbehavior, than controls perception and that simultaneously produces behavior asan unintended result. In that light, either of the alternative renderings yousuggested says something different from Bill P's original statement --Behavior: (is) the control of perception. Behavior is the means by whichperception is controlled; behavior is not a result of something, other thanbehavior, by means of which perception is controlled.
Of course, we then go on to say that many of the outward appearances ofspecific _actions_ we see a person make at a particular time are probablyunintended side effects of the person's control of perception. In that case, weare emphasizing the fact that, while behavior controls perception, the specificactions required to establish and maintain control must vary, moment-by-moment,any way necessary to eliminate the effects of disturbances that act on theperson's controlled variable(s). When actions vary, so do many side effects ofthe actions -- many of the "outward appearances" that often catch the eye of anuninformed observer, who then arrives at a wrong conclusion about what theperson is doing.
Specific actions are a result of any momentary changes in the person'sintentions (modeled as reference signals) _and_ in the effects of disturbancesthat affect the person's controlled variable(s). Those actions are the means bywhich the person controls perceptions.
If this isn't sufficiently confusing, just let me know. ;-)
Later, Tom
Date: Wed Jul 13, 1994 3:51 pm PST
Subject: Re: Diagram Terms o
[Paul George 940713 17:00]
{The original subject line was scrambled}
>[Bill Powers (940713.1230 MDT)]
> Below is our standard diagram of a control system with verbaldefinitions of the components and signals. Would you attach to this diagramthe labels you are accustomed to using for the parts of a controlsystem?
{"I reserve the right to revise and extend my remarks....."}
:
:
B ======
=========>|| F2 ||---C----------D
|| ====== |
^^ : v
|| : =====
====== : || F3 ||
A ===>|| F1 || : ======
====== : | E
^^ : | |
|| H ====== v ====== |
====<======|| F4 ||<-----G------ F <----|| F5||<---
====== ======
:
:
Note: Because of the way physical devices work, we can't (always) use justnegative feedback control. Two electric motors opposing each other like muscleswould burn out. Many devices like switches, pumps, and valves are discrete(on/off, forward/stop/reverse). And, as (I recall) Bill Leach observed, sensorsand actuators are expensive and so kept to a minimum. We also usually justsimulate continuous control because we are using electronics and digitalcomputers which (often) operate in a discrete fashion. Biological systems havecertain advantages.
Inside:
A = reference value or setpoint. Usually the desired state of H, butsometimes a limit - i.e. action should be taken when H exceeds the limit untilH falls below. Sometimes A = f(H), and corresponds to an estimated state of Eor F
B = Control output, output variable, or sometimes 'command'. Note that weusually keep it around and use it in computations (F1).
H = Control input, input variable, or input data (precise aren't we). Sameapparent semantics as in PCT controlled variable. In our case H can be a scalaror vector array, but that is an artifact of design. We could construct thingsyour way with one loop per variable.
F1 = Control logic or function Block. Usually not a simple, 2 variabledifference. Usually a number of H values and B values and (sometimes) A valuesgo through a set of algorithms to determine the desired value for B. We areusually forced to be pro-active due to latencies in the system and the processunder control.
{This _could_ have implications for HPCT at levels where there is aconsiderable delay between an action and any perceived 'result'. Back in schoolwe had some software in the Human Factors lab similar to the 'little man' wherethe user had to try to track and object with a cursor. The software insertedall kinds of linear and non-linear delays and vectors to the 'actions'. Madethe task real difficult, and frustrated people to no end}
F2 = Output point. An output transducer. Performs D to A conversion in somecases. Basically translates B to whatever the actuator instrument requires. Insome cases performs computations or sampling on B. Depends on the cycle time ofF1 vs that of the instrument.
F4 = Input Point. An input transducer. Usually performs A to D conversion.Usually samples or averages F
Note: Input and output points are referred to collectively as 'I/O Points'within the control system. They are often grouped into related sets (forpractical reasons) referred to as I/O Blocks, which are packaged as adevice.
Environment:
We would view the outside structure a little differently, because we viewthe Process (environment), instruments (nerves, muscles, eyes), and Controlsystem (brain) as being different things. In other words the system boundary isa little different. It doesn't really affect anything from the control system'spoint of view. I'll just re-diagram the input side
: |
====== ======= v ======
|| F4 ||<--G--|| F4.1 ||---- F <----|| F5 ||<---
====== ======= ======
:
F4.1 = Sensor Instrument. Translates the sensed variable F into some kindof a reading or metric. A pressure sensor turns a pressure into a voltage.Sometimes combines some set of physical sensors into a reading or set ofreadings. Note that I may measure the temperature of a vessel and translate itinto the temperature of the substance inside the vessel. Sometimes samples oraverages a continuous reading.
I think this has the same semantics as your G, but we consider it afunction, since it is a device. The channel is separate. Some advancedinstruments are control systems in their own right. They may have both sensorsand actuators, and may transmit an error signal.
The F4.1>F4 chain is analogous to the process of detecting light inseparate rods and cones, converting it to an optical nerve signal, and theninto an 'retinal image' (I'm fuzzy on the exact physical process). The imagethen needs to be processed to extract objects and motion which are 'perceived'.Some of this can map to control nodes, but having programmed such for radarsystems it is far from simple and involves a lot of computation andprediction.
G = Input signal or measurement. We would usually have a G' which is thechannel (wire), but the distinction is not important to the control system. Asnoted above, G in some cases is an error signal.
F - An attribute or characteristic of the process or of the equipment beingcontrolled . A temperature, pressure, etc.
F3, F5 = As in yours, though we would think of it as some characteristic orpart of the process system under control. We usually 'mirror' this functionwithin the control system for various practical reasons.
C - Could be a channel as in yours, but again the 'output' signal or'output data' is usually distinguished from the wire.
D - not real clear on its usage in PCT. The action of the actuator usuallyaffects the process or equipment being controlled, frequently indirectly
F2.1 Actuator Instrument. Translates a command into an action. E.g trips aswitch. Sometimes involves a sequence of actions which affect F. In a sense hassome of the semantics of D
Hope this hasn't munged things up too badly and has met your intent. Ithink this view can be translated readily into HPCT terminology as we appear tosimply be making slightly different groupings and distinctions. The only realdifference seems to be the definition of the 'Boundary' of the control systemand the 'Environment'. For simple PCT, environment appears to be everythingoutside of A, B, and H. For HPTC it is a point of view or (apparently)arbitrary scale {not a problem for any system modeler}. Process control justuses a kind of Mind/Body distinction that I haven't seen HPCT use explicitly.The distinction between a HPCT and multiple interacting PCSs or HPCSs isunclear, and probably unimportant.
The only likely difference I can see is that HPCT nodes can rout C of onenode directly into G of another without intermediary processing. In thedegenerate case B and H can be the same variable.
I am still thinking about how to directly use HPCT to modify processcontrol system architectures, but suspect I don't understand it well enoughyet. I think there is some real possibility of striking gold, but it is just anintuitive reaction. First theory, then application.
Later Paul
Date: Wed Jul 13, 1994 4:41 pm PST
Subject: Re: Behavior and control
[Paul George 940713 17:30] >Tom Bourbon [940713.1512]
> Behavior is the means by which perception is controlled; behavior isnot a result of something, other than behavior, by means of which perceptionis controlled.
> Of course, we then go on to say that many of the outward appearancesof specific _actions_ we see a person make at a particular time are probablyunintended side effects of the person's control of perception.
Could you clarify the distinction between Behavior and Action? I think itmay be the basis of much confusion.
I usually look at behavior as being a pattern or set of actions, usuallywithin the context of some environmental pattern of events or inputs. Trackingis behavior, a given arm motion is action.
You also seem to be saying that behavior directly causes other behavior.This makes sense in the context of muscles (unobservable) moving an arm(observable) or a arm motion propelling a ball. Is that all you meant?
Date: Wed Jul 13, 1994 4:42 pm PST
Subject: Re: Replies to Paul
From Tom Bourbon [940713.1538] >[Paul 940712 15:40] >>Tom
[940712.1341]
>> But what you have described, Paul, is an example of one controlsystem (A) disturbing a variable controlled by another (B) in such a way that Asees B's actions matching a pattern that A wants to see....
> Deliberately. I attempted to describe in PCT terms (or a reasonablefacsimile ;-) a useful idea from Skinner. Hoped I might demonstrate that Ihave some inkling of what you are talking about.
Good enough.
>> It is indeed true that Skinner described similar instances ofcontrol and counter-control, but he never understood why they worked as theydid -- he thought it was a matter of stimuli controlling actions. He missed bya mile.
> That he didn't 'properly understand' the mechanism is to me of littleimportance - that is the chaff. I'm just not a purist or true believer bynature.
But what you call "chaff" is the entire social and scientific establishmentcalled "radical behaviorism." The quality of Skinner's "understanding" is notsomething I dismiss as readily as you. He touted that understanding as _the_very science of behavior, not as a personal understanding. And from thatmistaken understanding (that behavior must be described as originating fromenvironmental stimuli), he drew up his guidelines for research, and forre-engineering society. His "chaff" is still used to justify operant techniquesand interpretations in many quarters of society. (Including the courts in theUnited States, where the idea that "the environment made me do it" is currentlyvery popular. However, in a demonstration of true eclecticism, the majorplayers in the same legal system appear to be enamored of the idea thatindividuals are "victims" of their own genes and brains. In our courts, you canhave it either way, or both ways at once -- either way, it's pure linealthinking -- all cause-->effect.)
> And as I commented to Bill P today, I think he meant that all youcould observe was stimuli apparently producing actions, and that wassufficient for behavioral modification.
I think you are right; I think Skinner _did_ mean that all _he_ couldobserve were stimuli apparently producing behavior. He went so far as to say(in so many words) that behaviorism is the science of that which can beobserved. But all that proves is that Skinner wasn't a very astute observer ofbehavior. Control _by_ organisms can be observed; it is an observable, not aconjecture. Skinner didn't observe it; therefore Skinner did not develop thescience of that which can be observed. Instead, he developed a technology,which he called a science, in which he treated all behavior as though it were aresult of environmental stimulation.
>>> A set of control nodes containing controlled variables producethe observed behavior, i.e it is the behavior of the system, not the parts.There _need_ not be a master variable. Distributed control systems do notnecessarily have to have 'master' nodes.
>> As Rick (asked), what does this have to do with a"gestalt?"
> We seem to have very different definitions for the term. I meansomething that does not appear until a 'critical mass' of components exist.Systems show group behavior, and it is sometimes a step function ofcomplexity. (N-1) components won't do it, and there must be the right Ncomponents. The little man doesn't work until all three control functionsexist, the sensory inputs exist for feedback, the output mechanisms exist,and all are properly interconnected.
I never thought of a control system as something that functions because thenumber of its parts is at least one greater than some threshold number ofparts. I think of a pile of parts as a pile of parts, not as a sub-thresholdsystem. And I always thought that systems "show system behavior," with eachsystem sort of "doing what comes naturally" for its particular organization. Onthat reading, a PCT system with missing parts isn't a PCT system withcritical-1 parts; it's simply not a PST system -- it is something else.
I'm still left wondering what this might have to do with "gestalt."
In these exchanges, we haven't even begun to explore the differencesbetween a hierarchical PCT model and the kinds of distributed control systemsto which you allude. In hierarchical PCT models, there are no distributednodes, each with its controlled variable(s). We seem to be talking aboutdifferent kinds of systems.
>> Paul, you are describing control as though it were a linealprocess of:
>> input--process ("determine")--select appropriate preprogrammedaction--act
>> I believe you left off that last step -- act.
> Nope, while action and input may be lineal, control is usuallycontinuous or at least periodic.
But you didn't describe it that way. You described a lineal process thatincludes the production of pre-programmed outputs, selected to matchpresent-time inputs. Or did I misread you then, and in this reply? It has beenour experience that lineal systems operating this way cannot control variablesin disturbance-filled world.
> . . I considered the 'selection' the same (from the controlsystem's point of view) as acting. The act is usually doing something to allowanother series of nodes to 'take action'.
In a reply to Bill Powers [Paul George 940712 17:00], you said more on thistopic:
> OTOH when I do process (a.k.a. activity) modeling, tasks havepreconditions that are evaluated to determine if the task needs to beperformed.Upon completion postconditions to affect the activation andprocessing of other tasks. The model tries to capture all potential tasks.Executing the model involves traversing appropriate paths.
I think these two replies (to Bill and me) are the clearest indications yetthat you are talking about a different kind of system from us. A system,incidentally, that is very similar to the ones described by Albus -- you mustthink of him as more than just someone to use as raw meat to cast before thePCT modelers. ;-)) You seem to envision a "system" as a kind of"meta-assemblage" of independent systems, each of which accomplishes anassigned task, then passes off the result to the next system(s) in the chain(or net) which uses the received result as fodder for its own processes, and soon. Then, in the language of Albus, the "behavior" of the entire assemblage"traces a trajectory" in space and time. I don't think that kind of system hasmuch in common with a hierarchical PCT model. Let us know what you think, afteryou have a chance to read some of our work.
By the way, we are acutely aware of the problem you mentioned -- part ofour work is published in our own somewhat inaccessible ghetto press -- but amight fine ghetto it is!
>> What's with this idea you keep going back to -- the idea thatthere must be more than one reference signal at a time. Of course that is thecase, but sometimes only one reference signal is needed in a working model.When more are needed, they are available.
> It had been stated as the general model, but every time an examplewas given it was in terms of 'a controlled variable' as if each loop had onlyone, usually refering to something complex. When I suggested a more complexset, I was told it was not needed. This produced confusion. I am relieved thatwe were saying the same thing (?) and that my initial understanding wascorrect.
I think Rick and Bill P. have had their shots at this topic with you, butlet me try.
At the risk of losing readers whose interests are limited to the seemingly"big" topics, I'll use the example of a person running a compensatory trackingtask. (No surprise there. I always hope that people interested in the bigquestions will come to realize this is not a trivially simple example of humanbehavior.) The person watches a computer screen and uses a control handle (h)to keep a cursor (c) aligned with a stationary target (t). A random disturbance(d) also affects the position of the cursor.
Paul, I know you will probably think this is a simplistic example, but Ihope you will bear with me; I am concerned over the fact that you and we seemto be talking about different kinds of models and that you express confusion atour bouncing you back and forth between talk about simple PCT models andcomplex ones. I hope to help clear up some of the problems.
The momentary position of the cursor is determined by two present-timevariables: the position of the handle and the magnitude of thedisturbance.
c = h + d
That equation completely describes the environmental variables in thetracking task. It is also the "environment equation" in the simplest PCT modelfor the person performing the task. Of course, the environment is really morecomplex than is shown in the equation: there is the room, the computer, thepower system, gravity, the air conditioning (the run occurs in Houston, in thesummer), perhaps other people, and so on. But for now we will use the simpleequation.
Now we know that the person is a rather complex device comprising what wecall organs, cells, systems, sub-systems, processes and so on. How will werepresent all of that in a PCT model? The simplest possible model includesthree functions (input, comparison and output) and three signals (perceptual,reference, and error). We assume the person establishes a reference signal (p*)(pronounced p-star) that specifies an intended relationship between cursor andtarget (c-t). In this case, let's assume
p* = 0, which means the intended difference of c-t = 0.
We assume the person compares p* with present-time perceptions (p) of therelationship, c-t. The comparison is a simple subtraction (p - p*) and thedifference is the error signal. e = p - p*
How can we plug that system into the loop with the environment equation? Weneed for the model of the person to do more than set p*, p and e; we need forit to behave. We assume that the person can be modeled as a negative feedbackcontrol system in which error signals drive the movements of its "handle." Howis error converted into movement, hence into handle position? In the simplestpossible PCT model, we use the equation (or program statement):
h = h - k(p - p*),
where h is momentary handle position and k is the integration factor, acoefficient which represents the movement of the handle (in units/sec/sec) forone unit of the error signal. That is the "person equation" in the PCT model.The complete model is:
c = h + d
h = h - k(p - p*)
That's it.
Now we know this is a pretty simple representation of a person and anenvironment, but it works. For tracking tasks, it duplicates the person'shandle movements and the resulting c-t relationships to a very high degree ofprecision -- correlations of .996 and higher between thousands of pairs ofmodeled and actual positions of the handle, and similarly for positions of thecursor. All other descriptive statistics -- and plain old eye-balling -- allconfirm the model duplicates and even predicts (as long as five years ahead oftime) the person's actions and their results.
With correlations of .996 and above, the simplest possible PCT modelaccounts for over 99% of the variance in the "simple" tracking task, but it canbe improved. For example, the model could be given an arm, like the one in BillPowers's "Little Man" program, and equipped with the Little Man's visualsystem. (By the way, "Little Man" is correct -- the modeled arm is programmedto resemble many physical features of Bill's arm.) We could explore many otherways to complicate the model. Some changes might degrade its performance,others might improve it. If they did improve it, the improvement would be inchipping away at the 1% or less of the variance that is not accounted for bythe simplest possible model.
So, we say the simplest PCT model is often good enough, but that we arewell aware it is inadequate as a complete model of the entire person and thecomplex environment.
I hope this treatise will help clear up some of your confusion over our wayof talking.
> Hope all these replies aren't eating up too much bandwidth.
Isn't that what bandwidth is for? Munch away! ;-)
Later, Tom
Date: Wed Jul 13, 1994 7:25 pm PST
Subject: $64,000 Question
[From Rick Marken (940713.1550)]
Re: Bill Power's diagram:
:
:
B ======
=========>|| F2 ||---C----------D
|| ====== |
^^ : v
|| : =====
====== : || F3 ||
A ===>|| F1 || : ======
====== : | E
^^ : | |
|| H ====== v ====== |
====<======|| F4 ||<-----G------ F <----|| F5||<---
====== ======
:
:
Paul George (940713 17:00) provides the following identifications
> A = reference value or setpoint.
> B = Control output, output variable, or sometimes 'command'.
> H = Control input, input variable, or input data.
> F1 = Control logic or function Block.
> F2 = Output point.
> F4 = Input Point.
> F4.1 = Sensor Instrument.
> G = Input signal or measurement.
> C - Could be a channel
> F2.1 Actuator Instrument.
Thanks. That helps. But I have a few questions.
1) I don't see the term "control variable" in your list. Wasn't it shown inthe diagram? What did you mean by "control variable"; what did you think Imeant by "controlled variable"?
2) I didn't find a definition of E. Don't you have a name for it? I wouldthink that it would be a very important variable in control engineering.
3) Here's the $64,000 question: What variable(s) is controlled in thisloop? Please explain your answer.
Thanks Rick
Date: Wed Jul 13, 1994 8:24 pm PST
Subject: Re: Diagram Terms o
<[Bill Leach 940713.22:11 EST(EDT)] >[Paul George 94071317:00]
> Note: Because of the way physical devices work, we can't (always) usejust negative feedback control. Two electric motors opposing each other likemuscles would burn out.
Which is a significant statement. From PCT one can rather easily concludethat when two people (two independent control systems) attempt to control thesame perceptual signals to different reference values, conflict will occur andone of the control systems might even "burn out" or at least "reorganize" untilthe error level is within acceptable limits.
In terms of engineered control systems, control loops may have positivefeedback paths, "feed-forward" and the like but if the overall control loop isever net positive feedback then you have what Dr. Armstrong used to make Radioreally practical... an oscillator.
> Many devices like switches, pumps, and valves are discrete (on/off,forward/stop/reverse). And, as (I recall) Bill Leach observed, sensors andactuators are expensive and so kept to a minimum. We also usually justsimulate continuous control because we are using electronics and digitalcomputers which (often) operate in a discrete fashion. Biological systems havecertain advantages.
Muscles also have "discrete" operation I believe. A particular cell eitheris or is not exerting force and the amount of force that a particular cellexerts when activated is not an available control function (though the numberof cells activated is and for how long or how often is).
Digital computers always operate in "discrete fashion". Even the famous D/Aconverter is a quantized device. OTOH, what we call analog electronics isconsidered to be continuous over whatever its operating range may be (that isit is not considered quantized unless one is talking about individualelectrons. Even there I don't believe that any theory maintains that the energyof a free electron is quantized).
> Inside:
> A = reference value or setpoint. Usually the desired state of H, butsometimes a limit - i.e. action should be taken when H exceeds the limit untilH falls below. Sometimes A = f(H), and corresponds to an estimated state of Eor F
The "but sometimes a limit" IS a reference. A control system does not haveto be continuously driving some perception to an exact value. For instance,often we will allow ourselves to get rather cool before we "reach a referencelimit" that results in our perceiving ourselves as feeling warmer (and maybeputting on a sweater).
> B = Control output, output variable, or sometimes 'command'. Notethat we usually keep it around and use it in computations (F1).
I don't believe that human physiology indicates any sort of "direct path"between perceptual sensors associated with output devices (ie: muscle tensionsensors) but there is clearly such control loops in existence. It is veryprobably (I think) that when we consciously perceive a physical controlproblem, it is likely that the perception comes from error signals "workingtheir way up the chain" or from other sensors such as the trembling and shakingthat can result when one stresses muscles for an extended period oftime.
For many, as I understand it, demonstrable reasons, it is highly unlikelythat perceptual sensors of output functions are used in high level controlloops such as is done in engineered systems. One obvious reason is the responsespeed differences that are often measurable.
> F1 = Control logic or function Block. Usually not a simple, 2variable difference. Usually a number of H values and B ...
This is not a problem for PCT on the conceptual level. That is, mostbehavior that we are likely to talk about ALWAYS involves a large number ofreference signals and their associated "control logic". However, from a humanphysiology standpoint, it is quite likely that a single "control block" willnever have more than one reference nor more than one perceptual signal at anyone time. It may have a number of "control" or "bias" lines connectedhowever.
The point here is that a series of control loops all connected together tocontrol a single high level reference can be discussed as though it were asingle monolithic control loop. It is valid to do so as long as one recognizesthat it is really multiple loops. In the example often given of someone pushingagainst someone else's hand while that second person is trying to control thepositions of their hand, the number of actual control loops involved isstaggering.
> Environment:
I believe that there is a serious "mangling" of control system theory goingon there. While in engineered systems, much of the environment IS"controllable" in the design, it is only controllable to an extent. This is thewhole reasons why engineered control systems exist... to produce consistentresults (goals) under varying and even unplanned environmental conditions(disturbances).
The mechanics of the design (digital computer with A/D converters, AnalogOperational amplifiers or an old steam turbine's flyball governor) are onlyimportant to the specifics of a particular application. They are all doing (orattempting to do) the same thing functionally. Again, they are attempting tokeep some measured parameter (perception) at a desired value (reference) byacting on the environment. It is the recognition that the Steam Turbine'sflyball governor is functionally the same as a billion logic gate controlsystem that is important.
-bill
Date: Wed Jul 13, 1994 9:03 pm PST
Subject: Re: Diagram Terms o
[Avery Andrews 940714.1400] (Bill Leach 940713.22:11 EST(EDT))
> I don't believe that human physiology indicates any sort of "directpath" between perceptual sensors associated with output devices (ie: muscletension sensors)
?? My recollection is that the spinal reflex loops involve two suchconnections, from organs sensing muscle tension near the tendons, and also fromspindles, roughly measuring muscle length. (I used to actually know this stufflast year, I'm horrified to discover how much I've forgotten in the last eightmonths, but there are plenty of books floating around in which it is all laidout). Or maybe I'm just misunderstanding something, such as what a `directpath' is supposed to be?
Avery.Andrews@anu.edu.au
Date: Wed Jul 13, 1994 10:19 pm PST
Subject: Re: Diagram Terms o
<[Bill Leach 940714.00:31 EST(EDT)] >[Avery Andrews940714.1400]
Oh hell! I thought that maybe you "miss quoted" my message (by leaving outpart of the statement) but then I went back and read the posting!
That should have read something along the lines of:
I don't believe that human physiology indicates any sort of "direct path"between perceptual sensors associated with output devices (ie: muscle tensionsensors) -- and the high level perceptual control functions.
What I was trying to say is that measurement of output function operationis likely not used by high level control functions directly. Normally,(always?) only if an uncorrected error exists in a lower level control loopwill a higher level control system "be concerned" with measured (perceived)output function performance (if then).
The so called "reflex loops" are complete control loops of their own (withtheir reference coming in as an output signal from a higher level controlsystem.
> Or maybe I'm just misunderstanding something, such as what a 'directpath' is supposed to be?
The reference to "direct path' was to compare the concept of how engineeredcontrol systems often actually sense an measure the operation of an outputdevice for use in calculation of system output signal values. I believe thatsuch functionality does exist in living control systems but that it isdistributed in the same fashion as the control hierarchy. That is muscletension sensors are generally inputs to a control loop directly associated withthe operation of the particular muscle (or maybe group) but are not an input tothe "top" of the control loop that has for its reference the act of picking upa book (for example).
Obviously I "blew it" in trying to express that idea so thanks for thechallenge to what I had actually stated.
-bill
Date: Thu Jul 14, 1994 6:52 am PST
Subject: Re: Facts, Theories & Illusions
[Dan Miller (940714)] Rick Marken:
Happy Bastille Day! - a day marked by good intentions, but the jail wasvirtually empty. Still, we should celebrate revolutions, and maybe we can bringabout our own.
Rick, thanks for the post. I don't really disagree with most of it. Ihaven't been very clear myself when I have discussed facts. Let me tryagain.
As I said, I see facts as perceptions that have been made significant. Thatis, the perceiver has detached the perceptions from the situation (temporallyand, perhaps, spatially). S/he speaks of them, thinks about them, tries to makesense of them, shows them to others (thus reproducible), and if lucky getsothers to agree about those perceptions. In this sense facts are symbolic andsocial objects (as in the object of attention and action). Also, facts assignificant social objects are used to support theories, models, hypotheses,hunches, etc. - or to negate others.
What I want to say is that the act of transforming the perceptions intofacts is itself based on some model or hypothesis about what the facts mean (orhow the facts relate to other observations). So, not only do we use theories tointerpret facts, but we also use some underlying theory to see and make senseof perceptions as facts. Somewhere in the higher reaches of the hierarchy(language and above) we organize, objectify, and relate our perceptions(sensations, intensities, durations, etc.).
You note (continually) that control is a fact. I agree. I am trying tofigure our what a fact is, how we discuss them, and how we use them toconstruct theories, destruct theories, (destruct? - sorry, I must have had someserious interference} that is destroy (or negate) theories, and a wholeassorted set of problems centering around shared perceptions and concertedsocial action. I am afraid this post is degenerating into drivel.
More later, Dan Miller
Date: Thu Jul 14, 1994 6:53 am PST
Subject: Re: Facts
[Dan Miller (940714)] Bill Powers:
Thanks for the post. I just replied to Rick Marken about what I am thinkingabout facts. Any comments are greatly appreciated.
I agree that the job of experimental scientists is to get as close aspossible to the bare facts (objective descriptions of perceptions?). However, Iam not so sure how close we can get to this sense data as humans. My thinkingis that those upper levels of the hierarchy really get in the way. Also theyallow us to do this kind of work.
Lots of others, particularly behaviorists and new cognitive scientists, donot see our facts as we do. They don't get their significance. My excursioninto this topic is as much about how they can be so blind to what are obviouslyvery elementary facts to us.
I'm on-line and getting interference, so I'll post more later.
Dan Miller
Date: Thu Jul 14, 1994 12:53 pm PST
Subject: Re: Replies to Paul
[Paul George 940714 11:00] >From Tom Bourbon [940713.1538]
> But what you call "chaff" is the entire social and scientificestablishment called "radical behaviorism." The quality of Skinner's"understanding" is not something I dismiss as readily as you. He touted thatunderstanding as _the_ very science of behavior, not as a personalunderstanding.
That is what I call 'fluffing the ware's' (legal & economic phrase).People, particularly scientists, scholars, and methodologists control forrecognition or status (I hope I phrase that correctly). This often leads toadvertising for product differentiation and declaring that one has the one truesolution. I concur with your conclusions about the effect of radicalbehaviorism.
> Instead, he developed a technology, which he called a science, inwhich he treated all behavior as though it were a result of environmentalstimulation.
Just for clarification, the distinction between technology and scienceis...??
There is a bit of a point of view problem. Clearly the environmentalstimuli (though that is a loaded term, I prefer 'influence' or 'event') affectsbehavior because stimuli is what can be perceived. When I have a continuous orrapidly cycling system (i.e. environment+organism) with what I deem cause andwhat I deem effect is largely an artifact of at what point I arbitrarily startto follow the chain of interaction. It is a classic chicken and egg situation.As I said earlier, I think PCTrs look at the situation from the organism'spoint of view while others look at from a theoretical external observer's pointof view. However, as Hiesenburg pointed out, the observer and observed affectone another whether they want to or not. PCT thus gives a clearer view of whatis going on as it recognizes interacting control systems.
> I never thought of a control system as something that functionsbecause the number of its parts is at least one greater than some thresholdnumber of parts.
I believe it has been demonstrated that there is behavior whose complexityis a step function of the number of components. It is certainly the case thatcertain concepts or theories cannot be understood until one has acquired acertain minimal set of facts (or other theories) - the pieces just won't fittogether. I can't control a process (a.k.a control a perceptual pattern) untilI have the right set of perceptions and actions to control it. I can't trackthe cursor or target unless I can recognize azimuth and elevation. Azimuth andrange won't cut it. I also can't perform the tracking task unless I have astick and the ability to move it in 2 directions.
Perhaps you should give me your definition of 'gestalt'.
> In hierarchical PCT models, there are no distributed nodes, each withits controlled variable(s). We seem to be talking about different kinds ofsystems.
Say what??? Then how would you describe a HPCS?? Must all of a controlnetwork occupy the same locus? Sight and motor control are the same parts ofthe brain and use the same variables? A HPCS can't involve more than oneindividual? I think we may be using different meanings for'distributed'.
Me
>> Nope, while action and input may be lineal, control is usuallycontinuous or at least periodic.
> But you didn't describe it that way. You described a lineal processthat includes the production of pre-programmed outputs, selected to matchpresent-time inputs.
Unfortunately language is lineal. Inputs arrive relatively constantly,evaluation occurs relatively constantly, and a given 'action programs' runconcurantly, though usually for a period of time. When I decide to throw apunch or execute a block, I do not have to concentrate on the micro movements.That task is spawned into to what is probably an autonomous control loop or'sub process'. It is a predefined pattern of actions that may be activated atneed. I am not constantly prepared to punch or block at all times. I have alittle trouble with the idea of a multitude of control loops monitoring at alltimes every existing controlled variable waiting for an positive error signalto occur. A cascade system seems much simpler and more efficient, but I amcontaminated by my experience as an engineer.
I guess the empirical question is whether all action in an organism is theresult of a continuously active control loop, or whether some are cyclical oractivated on demand (i.e. triggered by a discontinuous event). Note that if theperiod of a cycle is sufficiently rapid relative to the rate of change ofinputs and outputs it appears continuous. This is why we are able to usedigital computers for control.
Tracking a cursor, as you eloquently explained below, certainly is asituation where continuous negative feedback loop applies. You are in factcontinuously controlling something. It is not clear in my mind how or why theconcept of triggering a learned action pattern that goes away when the task iscompleted is inappropriate. Perhaps an extension of the 'little man' problem totwo little men playing the computer game 'pong' would be enlightening. Theproblem is probably my lack of understanding of HPCT.
> I think these two replies (to Bill and me) are the clearestindications yet that you are talking about a different kind of system from us.A system, incidentally, that is very similar to the ones described by Albus --you must think of him as more than just someone to use as raw meat to castbefore the PCT modelers. ;-)) You seem to envision a "system" as a kind of"meta-assemblage" of independent systems, each of which accomplishes anassigned task, then passes off the result to the next system(s) in the chain(or net) which uses the received result as fodder for its own processes, andso on.
Sorry for the confusion, I use process in two contexts. The first is wherewe are trying to provide a control system for a physical process, like a paperplant. The second is when I am modeling human processes for the purposes ofprocess improvement (SEI CMM or BPR or TQM style). It also applies forscheduling and planning projects. It could apply to the situation whereorganisms plan a future activity, like 'controlling for hunger' triggering ahunting behavior that involves a lot of other behavior patterns which areactivated by the recognition of certain environmental conditions along theway.
Your description of a system I think accurately describes both humansystems like companies, teams, and societies. It also applies to factories,herds of animals, and the 'environment' in general. It may apply to the controlsystems within the skin of an organism. The model pre-dated my reading ofAlbus. As I and (I think it was Rick) commented, Albus' model hadcorrespondence with real systems and a passing resemblance to HPCT structures.We may be talking about different kinds of systems, but I think they can beblended either by combination or translation. I'll let you know what I thinkwhen I understand more about the more complex PCT applications andconcepts.
Once more around the loop :-)
Paul
Date: Thu Jul 14, 1994 1:24 pm PST
Subject: PCT: A New Way to Look at Behavior
[From Rick Marken (940714.1045)] Paul George (940713 10:00)
> Has anyone in the CSG taken a close look recently on the work onArtificial Life, Core Wars, or Genetic Algorithm software?
I'm on the Santa Fe Institute's Artificial Life mailing list. The folks inSanta Fe are comfortably (and profitably, apparently) ignorant of the nature ofbehavior as control. Like those in the rest of the life sciences, they are busytrying to explain how organisms generate output (even if they call theseoutputs "goals"), not how they control.
Dan Miller (940714)--
> Lots of others, particularly behaviorists and new cognitivescientists, do not see our facts as we do. They don't get theirsignificance.
This is an excellent point. And I think I see what you mean about "facts"and I agree with you. From a PCT perspective, any fact (perception) isexperienced at many levels simultaneously. The facts we deal with in PCT haveto do with perceptions of "behavior" (whatever that is). We experience behaviorat many different perceptual levels at once; in terms of intensities,sensations, transitions (actions), events, etc (you know the drill). So a ratpressing a bar in a Skinner box is perceived as many simultaneous perceptual"facts"; color and shape of the rat, the relationship between paw and lever,the pressing event, etc. How one perceives these "facts" must depend on theperceptual functions that have developed in one's brain. Skinner developedperceptual functions that produced a perception of the environment shapingbehavior; he probably actually perceived (at some level) food pellets selectingor strengthening the behavior of the rat; "reinforcement" was a "fact" forSkinner as much as "control of reinforcement" is a fact for me. He SAW the ratbeing controlled by the reinforcement; I SEE the rat controlling thereinforcement.
Skinner would probably agree with a PCTer about the lower level facts ofwhat is going on in operant conditioning -- that the rat is white and that it'spressing the bar with its paw and that this results in the bar being pressedand a food pellet appearing. But Skinner's higher level perception of thesituation would differ from that of a PCTer; the PCTer would experience the"behavior" that Skinner sees as "pellet selecting bar press" as "rat selectingfood pellet".
So I think you make a very good point Dan -- "facts" (as higher levelperceptions) must, to a large extent, depend on the perceptual functions peoplehave built up and use to experience these facts. The perception of the samebehavior that Skinner has and that the PCTer has are still "facts" but I thinkone fact is "better" than another. The PCT "fact" is consistent with otherperceptions (like disturbance resistance) while the Skinner "fact" isnot.
This provides a nice sequel into my proposal for a new agenda for PCT. Ithink we should stop pushing PCT as a new theory of behavior because everyoneassumes that they know what the "behavior" is that PCT is a new theory of --and they are virtually always wrong. I suggest that we spend more timepublicizing PCT as a NEW WAY OF LOOKING AT THE FACTS OF BEHAVIOR. We have toteach people how to SEE CONTROL; we have to help people build new perceptualfunctions that will let them see what I call (perhaps incorrectly) the "fact"of control. In the operant situation, for example, we can do this by showinghow a perceptual aspect of the pellets (like rate in pellets/sec or amount involume/ sec) is maintained against disturbances (changes in schedule, size ofpellets, etc). This can be observed in real time if done properly. We have todevelop tons of demonstrations so that whenever a person sees a behavior, theyare able to see that there is a perception affected by actions and disturbancesand that the disturbances have no effect on the perception; we have to showthat this is happening when people push on buttons, type e-mail messages, setup psychology experiments, go to the library, brush their teeth, write aprogram, kiss their wife, pick up their daughter after work, make a barbecue,explain why alternative models of behavior are horseshit, eat lunch, make aphone call etc etc. People have to learn to perceive control. Until people cansee what we mean when we say that "behavior IS control" PCT will just beanother, rather unglamorous model of how people "behave".
I plan to suggest a number of "portable" demonstrations of control over thenext few days; I hope other people will have some suggestions for suchdemonstrations too. I'm sure there will be disagreements about whether what isbeing demonstrated is "really" control. But I think we should nail down thepresumed "facts" that PCT is designed to explain before getting too far intothe details of the PCT model itself.
I really think that the most important thing about PCT is that it gives usa new way of looking at what we call "behavior". Maybe a paper on "Watchingpeople control" would get the attention of behavioral scientists --maybe.
Oh, and we should also be clear that we know that not all behavior involvescontrol; sometimes people are flailing away, learning how to do something; inthis case the "something" being learned is not under control until the personlearns to control it.
Best Rick
Date: Fri Jul 15, 1994 2:28 pm PST
Subject: PCT and process control; modeling spatial perception
[From Bill Powers (940714.0900 MDT)] Paul George (940713.1700)
Excellent replies to my request for labels on the diagram. I'm going toconcentrate on main differences/agreements and not comment on side- issues,about which we have no _important_ disagreements.
:
:
B ======
=========>|| F2 ||---C----------D
|| ====== |
^^ : v
|| : =====
====== : || F3 ||
A ===>|| F1 || : ======
====== : | E
^^ : | |
|| H ====== v ====== |
====<======|| F4 ||<-----G------ F <----|| F5||<---
====== ======
:
:
> A = reference value or setpoint. Usually the desired state of H, butsometimes a limit - i.e. action should be taken when H exceeds the limit untilH falls below.
Right. We call A the reference _signal_, as it is variable. We call a"limit" system a "one-way" control system; if controls only when H exceeds (oris less than) A.
> B = Control output, output variable, or sometimes 'command'. Notethat we usually keep it around and use it in computations (F1).
This is what we call the error signal. It represents the difference betweenA and H. We wouldn't use the word "control" here; see below.
> H = Control input, input variable, or input data (precise aren't we).Same apparent semantics as in PCT controlled variable. In our case H can be ascalar or vector array, but that is an artifact of design. We could constructthings your way with one loop per variable.
Here there is some divergence. In PCT, H is the perceptual signal. It is_not_ what we call the controlled variable, because we use the term "variable"to refer to observables in the environment (what an external observer couldsee). A "signal" is a variable inside the control system. It is true that H isa controlled signal, in that the action of the whole loop is such as to make Hmatch A, whatever A may be. I don't much like "control input", because thisimplies that the signal H is capable of controlling something -- i.e.,maintaining it in a specified state.
Yes, you could construct things my way with one loop per variable; that'sthe current preferred model in PCT, but of course there's nothing carved instone there. Thinking of one loop per variable makes it easier to understandwhat a system is doing. Sometimes, I think, it makes a system easier to design,too.
> F1 = Control logic or function Block. Usually not a simple, 2variable difference. Usually a number of H values and B values and (sometimes)A values go through a set of algorithms to determine the desired value for B.We are usually forced to be pro-active due to latencies in the system and theprocess under control.
Again, I consider the term "control" unfortunate, because by itself thisblock can do no controlling. Control is what the whole loop does. We call F1the comparator, and treat it as a simple subtraction. This is part of seeinghow far we can go with the simple PCT model, as Tom Bourbon noted. Neurally,subtracting one signal from another to get a difference signal is very simple:one excitatory input and one inhibitory input do the trick. Of course thismeans that to get two-way subtraction you have to have two comparators, becauseneural signals can't reverse sign. We usually ignore this problem.
As to multiplicity of inputs to a comparator, I think that is taken care ofby the "one-variable-one-loop" architecture. Also I think that allowing toomany arbitrary connections in an ad-hoc way leads to messy designs, andprobably unlikely designs for a living system. And much of what you might tryto do with a complex comparator is taken care of in a simpler way by ahierarchy of control.
> F2 = Output point. An output transducer. Performs D to A conversionin some cases. Basically translates B to whatever the actuator instrumentrequires. In some cases performs computations or sampling on B. Depends on thecycle time of F1 vs that of the instrument.
Agree. We call F2 the Output Function. It converts the error signal intosome immediate influence on the physical environment, unidirectionally(disturbing its output does not alter its input).
> F4 = Input Point. An input transducer. Usually performs A to Dconversion. Usually samples or averages F.
We call F4 the Input Function or Perceptual Function. It converts one ormore inputs from variables like F into a signal H that represents some aspectof the part of the environment where F is. The kind of function performeddepends on the level in the hierarchy. For a system in the middle of thehierarchy, the inputs F are really copies of lower-level perceptual signals,some of which are under direct control and some of which simply representstates of the lower-level world. The first-level systems receive inputs onlyfrom the environment. Part of the modeling problem is to decide what functionF4 performs. It could be sampling, averaging, or anything else.
Since the output of F4, or H, is what is really controlled (the ultimatecriterion is that H must match A), it is the form of F4 that determines whatthe external observer will see being controlled in the environment of thecontrol system. Thus in PCT, we do NOT begin by knowing what aspect of theexternal world is under control, as the engineer would. We have to deduce whatis under control by hypothesizing forms of F4, then testing to see whether theenvironment, viewed through the hypothetical function, is resistant todisturbance because of the actions of the system. That's the basis for the Testfor the controlled variable. Engineers don't have to use that test; they knowwhat aspect of the environment is supposed to be under control.
> F4.1 = Sensor Instrument. Translates the sensed variable F into somekind of a reading or metric. A pressure sensor turns a pressure into a voltage.Sometimes combines some set of physical sensors into a reading or set ofreadings. Note that I may measure the temperature of a vessel and translate itinto the temperature of the substance inside the vessel. Sometimes samples oraverages a continuous reading. I think this has the same semantics as your G,but we consider it a function, since it is a device. The channel is separate.Some advanced instruments are control systems in their own right. They may haveboth sensors and actuators, and may transmit an error signal.
No problem here. In the nervous system, all input translations are eitherfrom a physical variable to a neural signal (considered as acontinuously-variable frequency at the lower levels) or from a signal into asignal. No A/D conversions as such, of course.
You are simply expressing the hierarchical nature of perception. At thelowest level, where stimulus intensity is converted to neural frequency, thereare some control systems (like the spinal reflexes or the iris reflex) whichdirectly control intensity signals, and thus indirectly control the physicalvariable (F) of which the signal is a quantitative analog. Copies of thecontrolled signal also pass upward to higher levels, the same levels from whichthe reference signals for the spinal control systems arise. So a higher systemmay send its error signal (B) to the reference inputs of a number of spinalcontrol systems, and receive its input information (F) in the form of copies ofthe controlled perceptual signals H from a number of intensity-control systems(as well as from uncontrolled sources). Thus the set of intensity-controlsystems becomes, on the output side, part of the output function (F2) of thehigher system, and also provides input data (F) to the input function (F4) ofthe higher system. The input function generally is a many-to-onefunction.
When we model just the higher system, we lump all the lower systems intothe higher system's output function, and treat the upgoing signals entering theinput function as if they were aspects of the environment. So all controlsystems at all levels operate through the environment, although at differentlevels of abstraction. We can experiment with higher levels of control withoutnecessarily understanding the lower levels that are involved.
So OK on your F4.1->F4 chain. In HPCT we actually have an F4.1...F4.11chain! With, of course, control systems at each level.
> G = Input signal or measurement. We would usually have a G' which isthe channel (wire), but the distinction is not important to the controlsystem. As noted above, G in some cases is an error signal.
I meant for G to refer to physical processes at too low a level to be ofinterest in modeling (like the jiggle of electrons in a wire). It's whateverconveys a physical quantity like temperature measured near the skin to thesurface of a cell like a temperature-sensor.
In PCT G would _never_ be properly called an error signal. The function F4might construct a perception representing the difference between twoenvironmental variables -- for example, the difference in position between afingertip and a target, as seen. But that is just another perceptual signal,representing the current state of some variable in the environment. Zerodifference in position might constitute an error, if the reference signal A isset to specify some non-zero difference that is to be maintained. In the LittleMan, non-zero reference signals for controlling target-finger separation in Xand Y are used to make the finger draw a circle around the target.
> F - An attribute or characteristic of the process or of the equipmentbeing controlled . A temperature, pressure, etc.
Yes. That, in HPCT, would be a first-level controlled variable, with Htstanding for the temperature, Hp for pressure, etc. With both T and P undercontrol, a higher-level system might receive the signals Hp and Ht as inputs ata second level, in which case those inputs become F for the higher system. Thenthe higher system could continually compute T/P, represent the ratio as asecond-level perceptual signal H2, and control the ratio of T to P (or anyother function of the two variables) by varying reference signals entering thetwo lower systems controlling T and P. Only the lowest level of system sendsits output directly into the environment via transducers. All higher systemsmust act by varying the reference signals for lower-level systems.
> F3, F5 = As in yours, though we would think of it as somecharacteristic or part of the process system under control. We usually'mirror' this function within the control system for various practicalreasons.
In PCT we distinguish these functions for a good reason. The function F3lies in the path between the visible physical output of the control system (D)and the visible controlled variable (F). The function F5 does not; it conveysthe effects of independent _disturbances_ (E) to the controlled variable by apath that is independent of the system's actions. In general, both E and theform of F5 are unknown to the control system and cannot be deduced from thebehavior of signals in the control loop. An example is the effect on your car'sdirection of a crosswind. The velocity of the crosswind would be E. Theaerodynamic laws that convert the vector velocity into a lateral force on thecar would be F5. You keep your car on the road without being able to sense Eand without knowing how to compute F5.
As you can no doubt see immediately, if F is under control, there will be anecessary relationship between D and E. This is the apparent relationship onwhich stimulus-response theory was built, with the existence of controlledvariables like F being unsuspected.
> C - Could be a channel as in yours, but again the 'output' signal or'output data' is usually distinguished from the wire.
OK. A detail.
> D - not real clear on its usage in PCT. The action of the actuatorusually affects the process or equipment being controlled, frequentlyindirectly.
Another critical point in PCT. D is the immediate effect of the actuator onthe environment. In a motor, it would be the torque generated by the motor onthe end of the shaft at the motor. This is rarely the variable that is to beput under control (F). What is usually to be controlled is some rather remoteeffect of D on the environment, such as the position of a load being wound upon a pulley at the other end of the shaft. The function F3 expresses thephysical link between the actuator output D and the variable to be controlled,F. Because of the function F3, it is possible for other influences in theenvironment to affect the controlled variable F independently of the output Dof the control system. The state of the controlled variable is really given byF = F3(D) + F5(E) (in the simplest case). This is why, if F is being activelycontrolled at zero, we get the interesting relationship that Rick Marken oftentalks about using other symbols: F3(D) = -F5(E). That's what gives theappearance of responses to stimuli.
The definition of D depends on the level of control you're talking about.It is always defined so that D is very difficult for the environment to affect,so D depends ONLY on the error signal in the control system.
F2.1: OK, just part of F3. Note that it's not part of F2 unless we knowthat the environment can't disturb it. The environment can actually insertdisturbances anywhere between D and F; we represent all disturbances, however,as an equivalent disturbance of the kind shown.
> Process control just uses a kind of Mind/Body distinction that Ihaven't seen HPCT use explicitly.
Right. The distinction in PCT is between the nervous system and all that isnot nervous system, with sensors and actuators lying in the boundary surface.This allows us to speak of "the control system" as an entity of relativelyfixed organization, while "the environment" can be continuously changing. Thesame control system can operate in many environments. This is seldom ofinterest in process control, because the engineered control system is bolteddown and wired to the process it is to control -- it can't wander around andencounter environments with different properties (F3, F5) the way mostorganisms can.
> The only likely difference I can see is that HPCT nodes can rout C ofone node directly into G of another without intermediary processing. In thedegenerate case B and H can be the same variable.
Right. Usually, however, we do use an output function (F2) even inhigher-level systems, because a given control system will act by adjustingreference signals for many systems of the next lower level. The leastcomputation we need is to set the signs of the outputs correctly so there isnegative feedback in the path involving each lower-level system. I have beentrying to avoid introducing complex output functions, because one of the niftyfeature of control systems is their ability to determine what is to becontrolled strictly in terms of the input function, F4. You don't need to getthe output function just so in order to get good control; it just has to havethe right sign and about the right form. Proportional or integral output willusually do, with perhaps some dynamic trimming. There's nothing fundamental tosay that output functions can't be complex, but my instincts tell me that weshould first explore what can be done with input functions, leaving complexoutput functions as a last resort. Some amazing things can be done with theoutput signals weighted only by 1, 0, or -1.
See my Byte articles, particularly Part 3:
Powers, W.T. (1979) The nature of robots: Part 3: A close look at humanbehavior. Byte,_4_, No. 8, 94-116.
> I am still thinking about how to directly use HPCT to modify processcontrol system architectures, but suspect I don't understand it well enoughyet. I think there is some real possibility of striking gold, but it is justan intuitive reaction.
Your intuition and mine agree.
Misc:
The _action_ of a system is D. D can in general affect many environmentalvariables in addition to F. So without finding which of these affectedvariables is under control, it can be difficult to decide what effect of Dshould be called the organism's "behavior." We mean by behavior the effect isthat is being controlled, not the action by which is it controlled. The action,D, varies as disturbances E come and go, and as internal properties of F3change. The real behavioral variable F changes only as the reference signal Achanges. This is why S-R laws are so unreliable and have to be deducedstatistically. They express relationships between D and E, but leave out theeffect of A: what the organism wants, which is variable.
What the organism is "really doing" can be seen only by looking at F. Ofcourse in another sense, you see what the organism is doing by looking at D.That's why we say that you can't tell what an organism is doing [F] just bylooking at what it's doing [D].
Best, Bill P.
Date: Thu Jul 14, 1994 1:37 pm PST
Subject: Re: Diagram labels
[Paul George 940714 14:30] >[Bill Powers (940714.0900 MDT)]
Thanks for a Very enlightening post. It cleared up a lot of minorconfusion. I agree that we don't seem to have any major differences.
The only real difference in terminology is kind of computer related. Foryou, and probably within living systems, signal and variable are equivalent.For us a variable is a static storage area whose value is changed by a signal.Anything which is interested in the signal reads the value of the variable. TheVariable represents or records the input or output. The distinction is notsignificant until you try to model a living system with a computer orelectronics
> In PCT G would _never_ be properly called an error signal. Thefunction F4 might construct a perception representing the difference betweentwo environmental variables
I meant to say, as you indicate above, that F might be a B of another loop.Sometimes it is the magnitude of the error signal that is of interest, not the'action' (C or D) which is the transduced output of F2. For simplicity in yourmodel you bundle a lot of transforms into F2 & F4 that we break out forpractical reasons.
> In PCT we distinguish these functions for a good reason. The functionF3 lies in the path between the visible physical output of the control system(D) and the visible controlled variable (F). The function F5 does not.
The distinction is very important, and critical for the _design_ of theprocess control system and the control algorithms. It is not important for theoperation of the system. Of course as you point out, an engineer knows the realprocess while an organism tries to divine it from its perceptions.
> I have been trying to avoid introducing complex output functions,because one of the nifty feature of control systems is their ability todetermine what is to be controlled strictly in terms of the input function,F4.... There's nothing fundamental to say that output functions can't becomplex, but my instincts tell me that we should first explore what can bedone with input functions, leaving complex output functions as a last resort.Some amazing things can be done with the output signals weighted only by 1, 0,or -1.
Ack, but it might make a more complete explanation of something likedriving a car or performing a learned response simpler. Starting your way isprobably best for Rick Marken who is doing low level modeling of physiology (itappears). However the complex output might be more useful for Dag in applyingHPCT to individuals or organizations.
Thanks again for the clarification, particularly the 'Misc'. A really slickobservation.
Paul George
Date: Thu Jul 14, 1994 1:41 pm PST
Subject: Re: $64,000 Question
Paul George 940714, 11:40 >[Rick Marken (940713.1550)]
> 1) I don't see the term "control variable" in your list. Wasn't itshown in the diagram? What did you mean by "control variable"; what did youthink I meant by "controlled variable"?
In our world 'control variable' means a data structure used in control,which contains a value. Thus it would apply to both B & H. The way I use itin this forum is to refer to H. It is that input that the control system wouldlike to get to match A through the change in (I guess) D.
> 2) I didn't find a definition of E. Don't you have a name forit?
Sorry, when I generated the post a second time (letting it simmer changedhow I structured my response) it got left off. It is some facet of the processthat we cannot directly sense, but must rather infer via F. I know of no formalname for it. Since we often cannot sense the perception we really want tocontrol, we often have to provide an F5 and/or F3 inverse within F4 or F1 (ormore usually between, this is the 'world model' or 'mirrored object'.
> 3) Here's the $64,000 question: What variable(s) is controlled inthis loop?
From your point of view (and mine) H. By convention and practice the systemis seen as controlling D in order to provide control of F (the 'process centricview'). However, the control system writ small (essentially F1) only caresabout H & B. Recall that in our system we are often triggering anindustrial machine (often with control capabilities) that may do a fairlycomplex series of things to the process under control. They respond to a fixedset of signal's that are usually interpreted as commands or instructions. An F3could represent such a machine.Sometimes B may be a program which is downloadedto replace an F3. Similarly a F5 could represent a machine not under the directcontrol of the control system (I know the phraseology flies in the face of theface of the PCT concept of what is controlled).
Hope this helps
Paul
Date: Thu Jul 14, 1994 1:42 pm PST
Subject: Re: Erratum
Paul George 940714 12:00 >[Bill Powers (940713.1700 MDT)]
> A comment on the environment part of the model is needed. In fact,the variable H might be multiple physical variables, H should be changed toD.
In that case there may be a disconnect if the statement is not true (for Dit certainly is).
In practice several Hs (and often B) must be input as some F1.n functionmust convert them to a value H' that may be compared to A. It is myunderstanding that allocating each H to a separate control loop usually doesn'twork because the value of a given A depends on the values of the other H's (notto say that it is absolutely impossible). See also my response to Rick's $64kquestion.
BTW It slipped my mind but F1, or more accurately {A,B,H,F1}, would oftenbe referred to as a 'Controller'. It could also refer to a HPCT network. It'sall a question of the physical architecture of the process controlsystem.
Paul
Date: Thu Jul 14, 1994 1:46 pm PST
Subject: Re: Diagram Terms o
[Paul George 940714 15:00] >[Bill Leach 940713.22:11 EST(EDT)]
> The "but sometimes a limit" IS a reference. A control system doesnot have to be continuously driving some perception to an exact value. Forinstance, often we will allow ourselves to get rather cool before we "reach areference limit" that results in our perceiving ourselves as feeling warmer(and maybe putting on a sweater).
What I meant is that there might several H's representing upper and lowercontrol limits SPC style. The error signal B could be {-1,0,1},or a larger set, resulting in different D's depending on which limit wascrossed. Slowing a car and slamming on the brakes can be a considered adifferent action, rather than just a difference in the gain of F2.
Paul
Date: Thu Jul 14, 1994 4:09 pm PST
Subject: We have a winner!
[From Rick Marken (940714.1445)]
Me:
> 3) Here's the $64,000 question: What variable(s) is controlled inthis loop?
Paul George (940714, 11:40) --
> From your point of view (and mine) H.
YES! You WIN!! Congratulations. You are the first control engineer on CSG-Lto correctly identify the perceptual signal as the variable (OK "signal") thatis controlled in a control loop! The behavior of a control loop is the controlof perception (seems like I've read that somewhere before).
A check for $64,000 will be in the mail to you soon -- I hope;-).
> By convention and practice the system is seen as controlling D inorder to provide control of F
Oops :-( The only variable in the loop that is controlled (kept at aspecified level against disturbance) is H. But you don't have to send the checkback (if you get it); getting H right is VERY impressive. Really. You wouldn'tbelieve how tough it is to get control engineers to accept this incrediblysimple fact.
> However, the control system writ small (essentially F1) only caresabout H & B.
Well, it really only cares about (that is, controls) H; the system careswhat value H is -- it "wants" H to equal A and it will do what it can to makethat happen. The system doesn't care what value B is; B just varies around asnecessary (depending on disturbances -- E -- and changes in functionalrelationships in the loop) to keep H matching A.
In PCT we like to say that a control system doesn't care what it does(variations in B and D), just what it perceives (variations in H).
Paul George (940714 11:00) --
> I believe it has been demonstrated that there is behavior whosecomplexity is a step function of the number of components.
What is behavior? What are its components? It sounds to me like thebehavior of which you speak consists of the visible actions and resultsproduced by an organism. Anything that has been "demonstrated" about thisbehavior would be quite irrelevant, would it not? After all, we know thatorganisms control their perceptions, right?
Best Rick
Date: Thu Jul 14, 1994 8:10 pm PST
Subject: Re: Behavior and control
From Tom Bourbon [940714.1740] Replies to Paul George and Bill
Leach.
>[Paul George 940713 17:30]
>>From Tom Bourbon [940713.1512]
>> Behavior is the means by which perception is controlled; behavioris not a result of something, other than behavior, by means of which perceptionis controlled.
>> Of course, we then go on to say that many of the outwardappearances of specific _actions_ we see a person make at a particular time areprobably unintended side effects of the person's control of perception.
> Could you clarify the distinction between Behavior and Action? Ithink it may be the basis of much confusion.
> I usually look at behavior as being a pattern or set of actions,usually within the context of some environmental pattern of events or inputs.Tracking is behavior, a given arm motion is action.
At least on first reading, that doesn't look terribly far from what Ibelieve is the PCT-modeling interpretation. I said not terribly far -- thereare some differences, the magnitudes of which we can explore. Most people,including most behavioral scientists, use the word "behavior" to refer to theirown observations of what another creature is doing -- he is walking, she istalking, they are assembling, the rats ran along that path to a new source offood, and so on. In that usage, behavior is interpreted in terms of theobserver's perceptual units and the labels are really names for _results_ theobserver notices when the other creature acts. In that usage, actions andbehavior are often equated or used interchangeably.
In PCT modeling, we have found it useful, often necessary, to think of thebehavior of a control system as what it is doing _from its own perspective_.If, after performing the Test for Controlled Variables, we concluded that asystem is controlling its perceptions of the position of an automobile on thehighway, we speak of that as its behavior -- that is "what it is doing." Itdoes what it is doing (which is controlling a particular perception in adesired way) by acting on the environment; the actions by which it achievescontrol of perception are not "what it is doing." The system's actions must be"out of control," in that the must vary any way necessary (for example, due toenvironmental disturbances) in order for the system to do what it is reallydoing -- controlling a particular perception or set of perceptions.
In the PCT interpretation, many, if not all, of the things an observer seesa system "doing" may well be outward appearances that are irrelevant to theobserved control system -- in most cases, what an observer sees does not evenexist for the observed control system -- it does not know that it is seen asdoing what the observer sees.
This difference in "points of view" concerning what the observed system isdoing is behind many misunderstandings between people -- "Why do you keep doingthat? Don't you know it drives me crazy?" "Stop doing what?" "Damn it, you knowwhat I mean! You keep making that stupid noise through your nose every time youtype at the word processor." "What noise?" And so on. And in the behavioral andcognitive sciences, we have scientists who observe people "doing things" andthen conclude that they (the scientists) know what those people are doing. Thenthey ask the people what they are doing. The observed persons reply from theirown perspectives; they describe what they are controlling, whereupon thescientists (and a few famous "neurophilosophers" I could name) shout inunison,"Aha! Do you see? We _told you_ that people are unconscious of what theyare doing. They don't really know what they are doing." In contrast to nearlyall other behavioral scientists (and those very famous neurophilosophers) a PCTmodeler is more likely to take people at their word when they say, "I wasmaking the contrast on the picture just right," or "I was going to the store tobuy flour to make cookies for Aunt Sally." The difference in our interpretationcomes from our knowing that the actions of the person are not "what the personis doing;" the person is "doing" his or her own controlled perceptions.
> You also seem to be saying that behavior directly causes otherbehavior. This makes sense in the context of muscles (unobservable) moving anarm (observable) or a arm motion propelling a ball. Is that all youmeant?
I don't think I was saying exactly that, but it is true that the actions ofa hierarchical control system are "nested" in something like the manner youdescribe. (As an aside, are you perhaps alluding to the old "behavior cannotcause behavior" song, from radical behaviorism?)
><[Bill Leach 940713.21:10 EST(EDT)]
>>From Tom Bourbon [940713.1512]
Bill L. says:
> Let me try to see if I can express this in a cogent fashion... Incommon terms, "behavior" is a label for actions of people. In its' commonusage, the term is pretty vague but that is how most people see the meaning.Behavior is what we observe another person do, that is their observed actionsare behavior.
Bill, in my reply to Paul I was actually replying to your questions aswell. Did I come close to addressing some of the questions you asked in yourpost? I agree that common usage is pretty vague on "behavior" and "actions" andeven more vague or silent on "unintended side effects of actions." In PCTmodeling, it is mandatory that we maintain some kind of clear distinction amongthe moment-by-moment products of the output function, the many environmentalconsequences of those outputs, and the perceptual signal(s) the system iscontrolling. In the modeling, the names of those variables are unimportant --they can be assigned any symbolic label that is acceptable to the programminglanguage one is using -- it is their _values_ that matter, when the PCT modelis run in simulations. Of course, once we leave the confines of the littleworlds we create when we are modeling, we are back in the world of people andwords; that's when the fun begins. This is why it is so crucial that anyone whowants to understand PCT look beyond the words they see on this net, or in oursometimes hard-to-locate publications. A deep understanding of PCT can onlycome from grasping the significance, if not the computations, of thequantitative modeling. The PCT model is not about the words; the words arealways inadequate for expressing the quantitative model, and the necessarilylinear structure of language can never convey the simultaneity and continuityof the model in action.
. . .
> A typical "formal" definition of the term "behavior" is: The term"behavior" may be understood to embrace both the expressed and potentialcapacity for activity in the physical, mental and social spheres oflife.
> <That looks pretty useless to me but that is what many will see asthe meaning of the term.>
> The phrase "Behavior is the control of perception." is a definitionof the term "behavior". I don't believe that I maintained an incorrectperception of the term as used in PCT but also recognize that most people willnot see that "behavior" is a output and nothing more (I say output because asI understand PCT, my "conclusions" when thinking are behavior even if they arenot observable to any outsider).
Do my comments above come close to addressing these ideas?
> It is also far more complete than what I said (this I see inretrospection). When I say something like "The control of perception resultsin what we call behavior." it is possible that the person hearing that mightthink that what is observed (so called behavior) is what is controlled (thoughI still think that such an interpretation is incorrect).
Agreed. You are speaking of the (easily understood by PCTers) problem anobserver has if the observer is not aware of the phenomenon of control. To thatobserver, it seems extraordinarily easy to identify what the observed system"is doing" and never even suspect that the conclusion is wrong. (An aside: Ibelieve this point is relevant to the current discussion on csg-l about"facts," but I won't have time to join in that discussion before I vanish forthe wedding.)
> I will have to think about this a bit more. I still feel like takingthe step to saying something like "The control of perception results in whatwe call behavior." and then going on to explain B:CP is not necessarily sucha bad idea.
I think you can see now why I said, in the earlier post, that some of uswould chose to disagree with you -- but good naturedly :-) -- if you were tosay that.
. . .
Later, Tom
Date: Thu Jul 14, 1994 11:51 pm PST
Subject: Re: Replies to Paul and Jeff
From Tom Bourbon [940714.1750] >[Paul 940714 11:00] >>Tom
[940713.1538]
>> But what you call "chaff" is the entire social and scientificestablishment called "radical behaviorism." The quality of Skinner's"understanding" is not something I dismiss as readily as you. He touted thatunderstanding as _the_ very science of behavior, not as a personalunderstanding.
> That is what I call 'fluffing the ware's' (legal & economicphrase). People, particularly scientists, scholars, and methodologists controlfor recognition or status (I hope I phrase that correctly). This often leadsto advertising for product differentiation and declaring that one has the onetrue solution. I concur with your conclusions about the effect of radicalbehaviorism.
It looks as though we ar approaching something of an agreement on Skinner,at least in terms of how he did his fluffing. That's good. :-)
>> Instead, he developed a technology, which he called a science, inwhich he treated all behavior as though it were a result of environmentalstimulation.
> Just for clarification, the distinction between technology andscience is...??
I was using Skinner's term. For some audiences, he said behaviorism wasonly a technology, not a science; for others, he claimed behaviorism was _the_science of behavior (TB: "the only" was understood). I assume that when he saidit was a technology, he was trying to evade challenges to the quality ofbehaviorism as science -- "You don't think behaviorism measures up as scienceand those are the reasons you give? Not to worry! Behaviorism is only atechnology -- a set of handy tools for controlling behavior." Notice the neatdance to the side? What a nifty way to stay out of arguments, one of thosemythical feats for which Skinner was noted during his life. Of course, the factthat there was more to behavior than radical behaviorists allowed to meet theireyes was ignored and behaviorists-as-THE-behavioral-scientists could go onabout their "Science" undisturbed. Come to think of it, Skinner was being adamned effective control system, wasn't he?
> There is a bit of a point of view problem. Clearly the environmentalstimuli (though that is a loaded term, I prefer 'influence' or 'event')affects behavior because stimuli is what can be perceived.
Agreed, up to a point. In the right conditions, it is possible to see acontingent temporal sequence of stimulus followed by response. (See Rick'spaper on the blind men and the elephant for a PCT analysis of that S-Rinterpretation -- oh, by the way Paul, Rick and the rest of us know the storywas around long before Rick used that as his title; that's why he picked thetitle.) But the fact that I can see that sequence does not mean I have seenwhat the observed system "is doing:" controlling specific perceptions. Once Isee _that_ fact, then an infinite array of specific stimulus-responsecontingencies can be predicted with great precision and they can be explainedand modeled as though the system acts to eliminate the effects of disturbancesthat affect variables related to its (the system's) controlledperceptions.
> When I have a continuous or rapidly cycling system (i.e.environment+organism) with what I deem cause and what I deem effect is largelyan artifact of at what point I arbitrarily start to follow the chain ofinteraction.
I guess that could occur, were you observing a truly discretesystem-environment interaction sequence. Can you identify some interactions inwhich that is really what occurs? Those aren't the kinds of things you see whencontrol systems interact with their environments, but there _are_ circumstanceswhen that can _appear to be_ the case. Again, see Rick's paper, and "Models andtheir worlds," by Bill Powers and me, and my paper on "Mimicry and repetition."Easy for me to say that : all three papers are in our ghetto journal, _ClosedLoop_. Mary Powers can supply copies, at cost.
> It is a classic chicken and egg situation.
Not in a closed loop.
> As I said earlier, I think PCTrs look at the situation from theorganism's point of view while others look at from a theoretical externalobserver's point of view. However, as Hiesenburg pointed out, the observer andobserved affect one another whether they want to or not. PCT thus gives aclearer view of what is going on as it recognizes interacting controlsystems.
Heisenburg wasn't talking about the behavior of organisms, so I'm notcertain we can assume that his comments will automatically apply to our work,but that caveat aside, I agree.
>> I never thought of a control system as something that functionsbecause the number of its parts is at least one greater than some thresholdnumber of parts.
> I believe it has been demonstrated that there is behavior whosecomplexity is a step function of the number of components. It is certainly thecase that certain concepts or theories cannot be understood until one hasacquired a certain minimal set of facts (or other theories) - the pieces justwon't fit together.
But PCT is about an empirical, nontheoretical phenomenon -- control, andabout the minimal organization of a system that can achieve control. It is notabout how many _pieces_ there must be, but about the specific functions thatmust occur, whether they occur in a person, or a pigeon, or a bacterium, or achemical reaction. A control system -- a person or a rat for example -- canloose many identifiable parts and still function as a control system. Whatmatters is not the absolute number of parts, but the specific organization ofthe ones that remain.
Incidentally, in Wales, Pedro Mendes gave an exciting presentation on hiswork in which he applies the PCT model to biochemical reactions. There is nonervous system involved there, or in the modeling of bacteria by Bill P. andRick. What is important is that the requisite functions occur. Their occurrenceis of course dependent upon enough "pieces" being present, whatever thesubstance of the pieces might be, but the number of pieces alone has nothing todo with what happens when they are connected and put in motion. It is thespecific organization of the system that matters. The organization representedin the PCT model seems to have wide applicability as a model for control -- weare willing to start suggesting that it will apply to control by all livingthings.
> I can't control a process (a.k.a control a perceptual pattern) untilI have the right set of perceptions and actions to control it. I can't trackthe cursor or target unless I can recognize azimuth and elevation. Azimuth andrange won't cut it. I also can't perform the tracking task unless I have astick and the ability to move it in 2 directions.
And having said all of those true things (ignoring for now the possibilitythat we might need to explore the meaning of "recognize azimuth and X"), you(generic you) have said exactly nothing about how it comes to pass that aperson can use a control stick to keep a cursor aligned with a target. Thatexplanation requires a model for the organization of the person, not for thenumber of parts in the person.
> Perhaps you should give me your definition of 'gestalt'.
Ah, it was _you_ who introduced the term into the discussion. I've called_your_ hand. Show your cards! :-)
>> In hierarchical PCT models, there are no distributed nodes, eachwith its controlled variable(s). We seem to be talking about different kindsof systems.
> Say what??? Then how would you describe a HPCS?? Must all of a controlnetwork occupy the same locus? Sight and motor control are the same parts ofthe brain and use the same variables? A HPCS can't involve more than oneindividual? I think we may be using different meanings for'distributed'.
See my post earlier today -- was it to Bill Leach? I was speaking of what Isaw (perhaps incorrectly) as an appeal by you to the kind of "distributedsystem" that is in Albus's papers. The PCT model, whether in the single-loop orhierarchical version, is not like the "models" in the Albus papers. You stillhave not told us whet your own model would look like, if not likeAlbus's.
>Me
>>> Nope, while action and input may be lineal, control is usuallycontinuous or at least periodic.
>> But you didn't describe it that way. You described a linealprocess that includes the production of pre-programmed outputs, selected tomatch present-time inputs.
> Unfortunately language is lineal.
Yes, it is. But you also included a diagram, and it was clearly a diagramof a lineal system for input-process-output.
> Inputs arrive relatively constantly, evaluation occurs relativelyconstantly, and a given 'action programs' run concurantly, though usually fora period of time.
But living control systems don't work "relatively" constantly; they workconstantly -- continuously. And the world works constantly, notintermittently.
> When I decide to throw a punch or execute a block, I do not have toconcentrate on the micro movements.
Yes. Isn't it amazing?! That's the nature of perceptual control -- theactions that produce your behavior -- your controlled perceptions -- seem to"just happen" as though by magic.
> That task is spawned into to what is probably an autonomous controlloop or 'sub process'.
Oops! I won't even start to reply to this until after you have a chance toread some things about PCT. If you still want to say this after you have donethe reading, then we can go at it in earnest. ;-)
> It is a predefined pattern of actions that may be activated atneed.
Ouch! In spite of what I said a sentence earlier, I can't hold back fromsaying that this is completely outside the understanding of behavior in PCT. Ibelieve we have developed some pretty convincing arguments that nothing likethis happens in control behavior. Nothing.
> I am not constantly prepared to punch or block at all times. I have alittle trouble with the idea of a multitude of control loops monitoring at alltimes every existing controlled variable waiting for an positive error signalto occur.
Read about it. I think you will enjoy it. And I think you will find thatthe characterization you just gave is not an accurate one of the hierarchicalPCT model. (That is not meant as a criticism or put down, but as anacknowledgement that you still have not read the literature or hung around thenet long enough to know what we say about such things.)
> A cascade system seems much simpler and more efficient, but I amcontaminated by my experience as an engineer.
That's OK. We'll forgive you for the bad company you fell in with, earlierin your life. ;-)
> I guess the empirical question is whether all action in an organismis the result of a continuously active control loop, or whether some arecyclical or activated on demand (i.e. triggered by a discontinuousevent).
The question, the $64,000,000 question (I've raised thee ten fold, Rick) isis there a "controlled variable." The next question is, "If there is acontrolled variable, what kind of system could control it?" The answer willalways be, "A control system could control it." And many other kinds of systems-- including all lineal systems -- could not.
> Note that if the period of a cycle is sufficiently rapid relative tothe rate of change of inputs and outputs it appears continuous. This is why weare able to use digital computers for control.
Fine, especially for people who want to make a living using digitalcomputers for achieving control. But that does not necessarily have anything todo with how living systems achieve control -- in fact, it probably has_nothing_ to do with how living systems control.
> Tracking a cursor, as you eloquently explained below, certainly is asituation where continuous negative feedback loop applies. You are in factcontinuously controlling something. It is not clear in my mind how or why theconcept of triggering a learned action pattern that goes away when the taskis completed is inappropriate.
Read "Models and their worlds," by Bill P and me. If after that you stillthink a triggered action pattern can explain tracking, I'll be prettydisappointed with our writing skills. (I'll stop short of saying, "I'll eat mycontrol handles.")
> Perhaps an extension of the 'little man' problem to two little menplaying the computer game 'pong' would be enlightening. The problem isprobably my lack of understanding of HPCT.
I don't have exactly that program, but I do have a lot of programs with twopeople, two hands, two models, and a model and a person, all interacting inabout five or six identifiable kinds of interactions. And I'm still overcomingmy profound ignorance about the amazing phenomenon of control.
>> I think these two replies (to Bill and me) are the clearestindications yet that you are talking about a different kind of system from us.A system, incidentally, that is very similar to the ones described by Albus --you must think of him as more than just someone to use as raw meat to castbefore the PCT modelers. ;-)) You seem to envision a "system" as a kind of"meta-assemblage" of independent systems, each of which accomplishes anassigned task, then passes off the result to the next system(s) in the chain(or net) which uses the received result as fodder for its own processes, and soon.
> Sorry for the confusion,
No apology needed. I'm just trying to understand your point of view -- andyou certainly are up against it, trying to understand ours.
Time to shut down and vanish for a protracted weekend, marrying off the"baby" in the family. If this post is filled with typos, I apologize -- it wasdone on the fly.
I can't imagine how I'll even read all of the mail I know will come in byMonday!
> Once more around the loop :-)
Keep chomping on that bandwidth!
Later, Tom
Date: Fri Jul 15, 1994 12:19 am PST
Subject: Re: Replies to Paul and Jeff
<[Bill Leach 940714.20:27 EST(EDT)] >Tom Bourbon[940714.0945]
> I didn't say what I wanted to say, or if I did it wasn't veryclear.
Not a problem. Indeed, I am almost aghast at the grammar and sometimesincomplete sentences that I seem to be using of late. I think that there isstill a serious "internal control conflict" present here. I have may thingsthat I should be doing rather than participating so heavily in the CSG-L (ofcourse, I would rather participate than not...).
Yes, I accept the idea that pulse rate is probably continuously variableover some range that represents the range of some signal and that such signalsare really analog not digital.
All the best for the wedding and see you when you "recover". :-)
-bill
Date: Fri Jul 15, 1994 12:55 am PST
Subject: Re: questions
[From Bill Powers (940714.1645 MDT)] Paul George (940714.1140)
I'm glad Rick asked about E, what we call "the disturbance." Thedisturbance represents the class of independent environmental variables thatcan act directly on the quantity you're trying to control, independently of theoutputs of your control system. Like someone pushing on your elbow while you'reaiming a pistol.
Your reply to Rick didn't clear up the point:
> It is some facet of the process that we cannot directly sense, butmust rather infer via F.
This would partly fit, except that in general there is no way to workbackward from the state of F (actually, H) to the cause of a disturbance, thatis, to E. All the control system knows is that H changed when the referencesignal A did not change. In a properly designed control system, the change willbe strongly resisted because of the closed-loop action. It isn't necessary toknow E. But a complete model of a control process has to include E; otherwiseyou'll be designing for a disturbance-free environment instead of a realone.
> I know of no formal name for it.
Don't tell me I have invented something. Actually, I doubt it. I learnedabout disturbances, after all, from reading control-system texts. Maybe thesubject has been dropped since the 1950s. If so, I might be dubious abouttrusting my safety to any control system designed since then.
One of the sources of misunderstanding we have had with control engineers(especially concerning "open-loop control") is that many of them seem toassume, in their designs, that all sources of disturbance can be accounted for,so they can be incorporated into a "world model." But that is a veryunrealistic approach to designing control systems, unless they are so simpleand transparent that you can anticipate everything the environment might do tothem. Well, that point aside, it is certainly not true of organisms that theycan identify all possible causes of disturbances and prepare to meet them. MOSTdisturbances are known only in terms of an unexpected and persistent change inthe controlled variable. It takes a closed-loop system to maintain control whenthat happens.
So am I to take it that in your process-control models, you do NOT make anexplicit provision for unpredictable disturbances to affect the controlledvariable?
> Since we often cannot sense the perception we really want to control,we often have to provide an F5 and/or F3 inverse within F4 or F1 (or moreusually between, this is the 'world model' or 'mirrored object'.
We never assume that either F3 or F5 (function linking a disturbingvariable to the controlled variable F) is known to the control system; theorganic system has to be able to maintain good control without knowing the formof either function.
It sounds odd to me to say that "we often cannot sense the perception wereally want to control." But then I realize that you mean we, the engineers,don't have available a sensor that can be put into the control system to let ITsense the variable we, the engineers, really want to control.
I should think, though, that it would often be possible to sense componentvariables and compute the state of the variable you really want to control.Isn't that one solution? That counts as sensing it, in PCT.
This is not a problem in modeling organisms. There is no engineer. If theorganism doesn't sense a variable, it can't be controlled by thatorganism.
> By convention and practice the system is seen as controlling D inorder to provide control of F (the 'process centric view').
Maybe I'll accept that. But I'd have to believe that the conventional viewis a hierarchical control view, which would surprise me. A more accurate way ofexpressing this idea as it is usually carried out is to say that the controlsystem VARIES D in order to control F. To control something means to bring itto a reference state and keep it there, in my dictionary. If you're driving acar, you can't simultaneously control the steering wheel angle and control thecar's position on the road. Not, that is, unless you mean this hierarchically:you send a VARYING reference signal to the wheel-angle control system, whichcontrols wheel angle to make it match the reference signal, and thus youcontrol the position of the car.
The reason you can't speak of controlling D is that D is determined just asmuch by disturbances E as it is by the reference signal A. You can't bring D toa predetermined state, because that may be the wrong state for counteracting adisturbance. The output action of the control system must NOT be controlled insome preferred state; it has to be free to vary as required to counteract theeffects of disturbances. Just imagine trying to drive while holding thesteering wheel at your favorite angle.
You can see why disturbances are considered important in PCT.
> Recall that in our system we are often triggering an industrialmachine (often with control capabilities) that may do a fairly complex seriesof things to the process under control. They respond to a fixed set ofsignal's that are usually interpreted as commands or instructions.
"Our system?" Are you speaking of the organism, or about the processcontrol systems you-all design? If you're speaking about the human system, Iwould dispute the claim that ANYTHING is done open-loop. There are alwaysdisturbances. But in industrial control systems, the system design is whateverthe engineers decide it should be. If they're confident that the process willalways produce exactly the result that the command specifies, without anyfeedback to check that what happened is what was wanted, and without any meansof altering the command if the process strays off track, then they are braversouls than I am. But it's their necks.
Don't tell me. Management says that sensors are too expensive.
I notice that NASA systems are quite variable in this regard. Sometimesthey are fully fed-back so that the result of every command, even everyswitch-transition, is immediately observed, and the command can be changed orcanceled if the wrong result even starts to occur (or if the contacts don'tclose). Others are designed by engineers who seem never to have heard the term"disturbance." When they send commands to those systems, Mission Control sendsan astronaut to stand by and observe the actual result, and report it backimmediately. Good thing, too.
> Sometimes B may be a program which is downloaded to replace an F3.Similarly a F5 could represent a machine not under the direct control of thecontrol system (I know the phraseology flies in the face of the PCT concept ofwhat is controlled).
I don't care about phraseology as long as I can figure out what you'retalking about. I suppose that a control system could be built that senses whichprogram is running, which program should be running, and if there is an errordownloads a replacement program to a lower-level system. So far we haven't doneany experiments complicated enough to call for that model. But there's a placereserved for it, at the program and principle levels in HPCT.
Actually, since engineers aren't constrained by any evolutionary orsurvival considerations, they can design systems any way they want, as long asthey do the job at hand. They aren't designing general-purpose control systemslike organisms. They don't have to worry about ad-hoc patches to an existingdesign in terms of how it is going to affect the whole system in a differentsituation. If it works, screw the cover down and ship it.
I've been intermittently seeing the TEMPUS downlink video from STS-65. Thisexperiment includes a levitation device for "containerless processing" of smallspherical alloy samples. Maybe there's some basic problem about stabilizing thesample spheres, but after five days of watching I'm going crazy from seeingthat sample jiggling and bouncing around with every little disturbance from thespacecraft, or from heating or from evaporation or from who knows what. Jiggle,jiggle, jiggle. At least three samples that I know of have hit the cage whilemolten and stuck to it, ruined. And then the PI says "Thanks, Columbia, it'snice and stable now." Jiggle, jiggle, jiggle. Jiggle, jiggle, jiggle. I'mprobably doing the designer a terrible injustice; the control problem might beall but unsolvable. But I want to grab the phone and yell at the PI, "For God'ssake, haven't you ever seen a REAL control system?" I have enough hubris tothink that if someone asked me to stabilize a levitated sphere, they'd think itwas nailed in place, like the parrot's feet. Sometimes being old and retired ishell.
Best, Bill P.
Date: Fri Jul 15, 1994 3:19 am PST
Subject: Re: Diagram Terms o
<[Bill Leach 940714.20:34 EST(EDT)] >>[Paul George 94071415:00]
Indeed, Paul, as you have mentioned (and others before you, includingmyself), how you choose to view a particular control system example influenceshow you might "parse" the parts.
In the simplest biological control loops, it appears that a referencesignal can only have magnitude (including zero) but may not change signs or "gonegative". However, it also appears that there are "complementary controlloops" where such bidirectional control is necessary (most all physicalmovement can be viewed that way but I am not sure that even that view is reallycorrect).
> Slowing a car and slamming on the brakes can be a considered adifferent action, rather than just a difference in the gain of F2.
Yes they could be but probably are not. Then again... there might actuallybe two types of situations. I am thinking of the person that appears to reallyhave two modes of operation of the brakes; Variable pressure control to slowthe vehicle at some controlled rate is one and the other is "panic" mode wherethey just "press for all that they are worth" (hopefully at least on the brakebut often not!).
-bill
Date: Fri Jul 15, 1994 8:15 am PST
Subject: Re: Behavior and control
[From Rick Marken (940715.0730)] Tom Bourbon (940714.1740)
Wow! What a nice post, Tom. You've managed to say, very clearly andconcisely, what I've been struggling to say all week. Here are some of myfavorites:
> Most people, including most behavioral scientists, use the word"behavior" to refer to their own observations of what another creature isdoing -- he is walking, she is talking, they are assembling, the rats ranalong that path to a new source of food, and so on. In that usage, behavioris interpreted in terms of the observer's perceptual units and the labels arereally names for _results_ the observer notices when the other creature acts.In that usage, actions and behavior are often equated or usedinterchangeably.
> In PCT modeling, we have found it useful, often necessary, to thinkof the behavior of a control system as what it is doing _from its ownperspective_...It does what it is doing (which is controlling a particularperception in a desired way) by acting on the environment; the actions bywhich it achieves control of perception are not "what it is doing." Thesystem's actions must be "out of control," in that the must vary any waynecessary (for example, due to environmental disturbances) in order for thesystem to do what it is really doing -- controlling a particular perception orset of perceptions.
> In the PCT interpretation, many, if not all, of the things anobserver sees a system "doing" may well be outward appearances that areirrelevant to the observed control system -- in most cases, what an observersees does not even exist for the observed control system -- it does not knowthat it is seen as doing what the observer sees.
> This difference in "points of view" concerning what the observedsystem is doing is behind many misunderstandings between people.
And the grand finale:
> The difference in our interpretation comes from our knowing that theactions of the person are not "what the person is doing;" the person is"doing" his or her own controlled perceptions.
Now THAT'S PCT!!!
You've done well, Tom. Now you can go off and enjoy your daughter'swedding; and don't worry about how much it costs; the grants will be pouring innow that you've explained with crystal clarity what PCT is about.
Mazel Tov Rick
Date: Fri Jul 15, 1994 11:11 am PST
Subject: Gestalt
[Paul George 940715 10:30] >From Tom Bourbon [940714.1750]
Me:
>> Perhaps you should give me your definition of 'gestalt'.
> Ah, it was _you_ who introduced the term into the discussion. I'vecalled _your_ hand. Show your cards! :-)
I've been trying. And from your discussion I think we are saying much thesame thing. One more time.
To perceive something which exists in the environment, one needs a set ofinputs. The blind men couldn't correctly perceive the elephant because theydidn't touch enough of it and couldn't see the whole. Similarly, I can't clapwith only one hand. It isn't the number of components (def differs betweenexamples) that matters, it's just that if you don't have the complete set,nothing happens. At some point you have enough knowledge, or control loops, orperception, or instrumentality to affect the environment in a way that allowsyou to control your perceptions effectively (hope that was bad enough wording:-). Adding more may increase your level (effectiveness) of control, or maynot. But less means you can't control at all. At some point there is a criticalmass.
A Baby can't act very successfully because it hasn't learned enough aboutwhat the sensory stimuli 'means'. I can't form complete sentences until I knowenough words and grammar. Children don't comprehend conservation of area orvolume until a certain age (exactly why is not apparently known). We maypresume their HPC network is not sufficiently complex. In general there is somecorrelation between the number of neurons & connections and 'levels ofintelligence' (speaking of fuzzy terms). I don't think it is a smooth curve(perhaps Rick can tell us).
Hope these examples clarify what I am talking about.
Paul
Date: Fri Jul 15, 1994 11:13 am PST
Subject: Chapman and Agre
[Paul George 940715 10:05]
Ok, I just gotta ask. {Putting on asbestos suit to avoid fire &brimstone}
What is it that the heretics :-) Chapman and Agre say that differs from'pure' PCT? They seem to catch a lot of flack in this forum.
Date: Fri Jul 15, 1994 12:39 pm PST
Subject: Re: Behavior and control
[Paul George 940715 09:30] >Tom Bourbon [940714.1740]
Me:
>> You also seem to be saying that behavior directly causes otherbehavior. This makes sense in the context of muscles (unobservable) moving anarm (observable) or a arm motion propelling a ball. Is that all youmeant?
> I don't think I was saying exactly that, but it is true that theactions of a hierarchical control system are "nested" in something like themanner you describe. (As an aside, are you perhaps alluding to the old"behavior cannot cause behavior" song, from radical behaviorism?)
I have no problem with behavior causing behavior (and don't remember thesong). I just wondered if you intended something more than the simpleinterpretation. However, given your definition of behavior above (what theobserver perceives) I'm no longer sure I understood what you meant.
Have fun at the wedding (Free at last! Free at last.... :-)
Paul
Date: Fri Jul 15, 1994 12:39 pm PST
Subject: Re: Diagram Terms o
[Paul George 940715 12:00] >[Bill Leach 940714.20:34)]
> Then again... there might actually be two types of situations. I amthinking of the person that appears to really have two modes of operation ofthe brakes; Variable pressure control to slow the vehicle at some controlledrate is one and the other is "panic" mode where they just "press for all thatthey are worth" (hopefully at least on the brake but often not!).
That is the situation I had in mind. People often react very differently togaining on another vehicle, and seeing a truck stopped in front of them.
Paul
Date: Fri Jul 15, 1994 2:27 pm PST
Subject: Misc
[From Dag Forssell (940715 0930)] >[Paul George 940708 15:00]
> Looking forward to next week.
It has been a good week. A handful more from BPR-L are lurking.
Oldtimers on CSG-L are committed to PCT and try to be very careful in theiranswers, no matter who asks; no matter what the question, because there aremany lurking, watching the discussion and learning from it. I learn too. Ithink it has been abundantly clear from discussions this past week that PCT isa serious physical science, not an "off the top of your head" wordmodel.
-------------------------------
I have found the discussion on FACTS very useful. Thanks, Dan.
Best, Dag
Date: Fri Jul 15, 1994 2:27 pm PST
Subject: Re: questions
[Paul George 940715 11:45] >[Bill Powers (940714.1645 MDT)]
> Don't tell me I have invented something. Actually, I doubt it. Ilearned about disturbances, after all, from reading control-system texts.Maybe the subject has been dropped since the 1950s. If so, I might be dubiousabout trusting my safety to any control system designed since then.
There probably is a formal name, the problem is that I am a softwareengineer by training, not a control engineer. I work on development processesand techniques not process control software. Unfortunately, most peopledeveloping control systems these days aren't trained in control theory. Mycompany used to have a kind of apprentice program for all engineers for thatpurpose, but it was dropped about 15 years ago. Fear and trembling isappropriate. Read the RISKS forum (ACM SIGSOFT Transactions or comp.risks. isalso a mailing list if you are interested).
> One of the sources of misunderstanding we have had with controlengineers (especially concerning "open-loop control") is that many of themseem to assume, in their designs, that all sources of disturbance can beaccounted for, so they can be incorporated into a "world model."... So am I totake it that in your process-control models, you do NOT make an explicitprovision for unpredictable disturbances to affect the controlledvariable?
Yup. We are forced to try to anticipate all possible problems and deal withthem. In fact for safety critical systems we are legally required to. That iswhy requirements analysis and design are so difficult. If somethingunanticipated does occur and something bad happens, we can be held legallyaccountable if we 'knew or should have known' that it could occur. The bigproblem is that if an E (disturbance) occurs we usually must have a D to affectit (making it another F). Disturbances (E) which may be corrected via D are nota problem. Our task is to insure we actually have all the Fs and D's that arerequired to keep the process under control. Fortunately the task is eased byhaving a human operator as part of the System. They are a bit more flexible anddon't just 'fly IFR' (instrument flight rules).
> But then I realize that you mean we, the engineers, don't haveavailable a sensor that can be put into the control system to let IT sense thevariable we, the engineers, really want to control.
Or more likely, the customer wouldn't pay for it, or there is no place toput it. Sensors and wiring are expensive and kept to a minimum. One of thereasons that continuous control isn't used is that message passing networks areused for communication between controllers and instruments to minimize cabling.Point to point wire connections, nerve style, would be prohibitively expensive.There are also transmission problems with analog control signals and longcables. Practicality raises its ugly head.
>> By convention and practice the system is seen as controlling D inorder to provide control of F (the 'process centric view').
> Maybe I'll accept that. But I'd have to believe that the conventionalview is a hierarchical control view, which would surprise me.
These days it is, at least in companies who build plant control systems.Control engineers working at the PLC loop level is probably what you arefamiliar with. Hierarchical systems emerged in the 70's. The car example isexactly how we usually design it.
> "Our system?" Are you speaking of the organism, or about the processcontrol systems you-all design?
My whole post dealt with process control systems, or more accuratelyhierarchical distributed plant control systems. Organisms work differently,though system architectures are patterned on various understandings of howorganisms function.
> Don't tell me Management says that sensors are too expensive. Inotice that NASA systems are quite variable in this regard.
NASA doesn't _sell_ control systems. Their contractors (system developers)are in the business of selling engineering hours. In the industrial world,system proposals are evaluated mostly on price. The number of I/O points andcable lengths are the major cost drivers.
> I suppose that a control system could be built that senses whichprogram is running, which program should be running, and if there is an errordownloads a replacement program to a lower-level system.
That is actually a normal situation. If a processor crashes, you often needto reload the software. But I meant something more akin to learning. Youeffectively replace the old control node (or HPCS) with a new one that isbetter for handling the current situation, or is more efficient (say adifferent F2 or F4). In addition, since (due to bad design) the H values areoften hard coded into the software, to change it you need to download a newversion.
Later Paul
Date: Fri Jul 15, 1994 5:38 pm PST
Subject: Chapman and Agre
[From Bill Powers (940715.1350 MDT)] Paul George (940715.1005)
> What is it that the heretics :-) Chapman and Agre say that differsfrom 'pure' PCT? They seem to catch a lot of flack in this forum.
Just from me, and you can put it down to prejudice. They seem to make up alot of peculiar terms for what turn out to be simple things, like theirproposal that behavior is "situated." This means that behavior takes place inan environment, as near as I can figure. For AI types this may be a majordiscovery, but not for anyone else. They're probably doing no harm; I justdon't go in for high-falutin' generalizations. There are lots of truestatements one can make, the only polite response to which is "My, that _is_interesting." I have a feeling that they belabor me with earnest urgings tograsp simple facts, as if they see something in them of, unfortunately,inexpressible importance.
So now you know that I'm not very impressed by their work. Won't I looksilly when they turn out to be right?
Best, Bill P.
Date: Fri Jul 15, 1994 6:53 pm PST
Subject: Re: We have a winner!
[Paul George 940715 10:00] >[Rick Marken (940714.1445)]
Me:
>> By convention and practice the system is seen as controlling D inorder to provide control of F
> Oops :-( The only variable in the loop that is controlled (kept at aspecified level against disturbance) is H.
Ack. I just meant that that is the way process engineers usually talk aboutit. And while PCT usually just works within the context of the control loop, Inprocess control, the Instruments and the equipment performing the process areviewed as part of the 'System'. The system boundary is in a slightly differentplace. Also from a terminology standpoint (backward from PCT) the thing actedupon is seen as being controlled, the sensors just tell you what happened (youhope :-) {sound familiar?}.
>> However, the control system writ small (essentially F1) only caresabout H & B.
> Well, it really only cares about (that is, controls) H; the systemcares what value H is -- it "wants" H to equal A and it will do what it can tomake that happen. The system doesn't care what value B is; B just variesaround as necessary (depending on disturbances -- E -- and changes infunctional relationships in the loop) to keep H matching A.
I understand that is the way it is usually modeled. But is it necessarilyimpossible that B is also 'perceived'? To anthromorphisize a bit, If I push ona door lightly and don't perceive it moving, I then push harder. To increasethe force I must have some idea how hard I was pushing originally. Now, in thisparticular case I have sensors that tell me how hard my hand is pressing, orhow hard my muscles are contracting. But the concept of direct feedback when Ican't detect (indirectly) C or D doesn't seem invalid on the face. The signalexists, and the control loop could 'wire itself' that way.
Date: Fri Jul 15, 1994 6:54 pm PST
Subject: Heretics, etc
[From Rick Marken (940715.1300)] Paul George (940715 10:05)
> What is it that the heretics :-) Chapman and Agre say that differsfrom 'pure' PCT?
Chapman and Agre are not heretics because they were never in the cult (ofPCT that is). Like most behavioral scientists I imagine that they are eitherblissfully unaware of PCT or have no idea what it is about.
> They seem to catch a lot of flack in this forum.
No more than any other psychologists who are clueless about the nature ofcontrol- - ie. all of them. They were once mentioned (long ago) as people whosework was (unbeknownst to us or them) compatible with PCT because they talkabout "situated action" and seem to be aware of the fact that actions have toadjust to environmental circumstances (disturbances) if people are to achievetheir ends. But, like all other psychologists, they have no idea that thismeans that it is perception, not observable action or behavior, that is"achieved". So they are busy building superficial models of superficialbehavior. And, of course, ignoring PCT.
The only PCT "heretics" I know of are Carver and Scheier and Hyland (thereare a few other names as well but I can't remember them). These people are"heretics" only because they talk in terms of Powers' PCT model and refer toPowers a lot but they are applying the PCT model to the wrong phenomenon -- toobservable actions and behavior (which are usually, as Tom noted, irrelevantside effects of control) rather than to control.
Paul George (940715 10:30)--
> In general there is some correlation between the number of neurons& connections and 'levels of intelligence' (speaking of fuzzy terms). Idon't think it is a smooth curve (perhaps Rick can tell us).
I'm not a physiologist; I study control at the same levels as Tom and BillP. ie. all levels. But I will say that I don't think that a correlation that isless that .99 is of much interest in PCT. What's your guess at the correlationbetween number of neurons & connections and 'levels ofintelligence'?
Rick
Date: Fri Jul 15, 1994 7:32 pm PST
Subject: Control of output? No such thing
[From Rick Marken (940715.1530)]
Me
>> Well, it really only cares about (that is, controls) H; the systemcares what value H is -- it "wants" H to equal A and it will do what it can tomake that happen. The system doesn't care what value B is; B just varies aroundas necessary (depending on disturbances -- E -- and changes in functionalrelationships in the loop) to keep H matching A.
Paul George (940715 10:00) --
> I understand that is the way it is usually modeled. But is itnecessarily impossible that B is also 'perceived'?
Why don't we go back to PCT terminology now. B is the error signal in acontrol loop. It would make no sense to have it be perceived by the controlloop itself (perhaps re-entering the loop through the perceptual function - -F4 in Bill's diagram, or added directly to the perceptual signal -- H in thediagram -- I don't know what you had in mind) though the error signal couldbecome the perceptual input to another control loop.
> To anthromorphisize a bit, If I push on a door lightly and don'tperceive it moving, I then push harder.
Right.
> To increase the force I must have some idea how hard I was pushingoriginally.
Not at all, if by "some idea" you mean that the control system must have orwould be helped by a perceptual representation of the force that is beingexerted on the door (variables C or D in the diagram). All the control loopneeds is a perceptual representation of the variable under control -- theposition of the door in this case -- which is variable F in Bill's diagram; itis called the controlled variable in PCT.
> Now, in this particular case I have sensors that tell me how hard myhand is pressing, or how hard my muscles are contracting.
You only need those sensors if you are controlling those variables. Itlooks like you've let psychologists sell you a bill of goods about how controlsystems work. A control system only needs to perceive the variable it iscontrolling; it does not need to (and typically CANNOT) perceive its ownactions (outputs) , the actions of the environment (disturbances) or how thoseactions (outputs and disturbances) affect the controlled variable.
> But the concept of direct feedback when I can't detect (indirectly) Cor D doesn't seem invalid on the face.
Just one of the many surprising results of analyzing the operation of acontrol system WITHOUT PRECONCEPTIONS;-)
A simple control loop can detect neither its own output (C in Bill'sdiagram) or the effect of this output on the controlled variable; nor can itdetect any disturbance(s), E, or the effect of any disturbance on thecontrolled variable.
Is PCT starting to smell a little revolutionary yet?
> The signal exists, and the control loop could 'wire itself' thatway.
But there would be absolutely no reason for a control system to wire itself"that way". I can't think of any way you could add a "perception" of thesystem's own error signal or output that wouldn't be either superfluous or amajor hinderance. Maybe you could draw a diagram of what you had inmind?
Best Rick
Date: Fri Jul 15, 1994 9:02 pm PST
Subject: Re: Diagram Terms o
<[Bill Leach 940715.22:32 EST(EDT)] >[Paul George 94071512:00]
> That is the situation I had in mind. People often react verydifferently to gaining on another vehicle, and seeing a truck ...
Yes but the difficulty is that this is something that we often observe anddirectly tells us nothing about what the person is controlling. We can probablysurmise that the control system is either making a discontinuous change inreference signals or a perception involved in a loop with exceptionally highgain suddenly deviated from its' reference.
Even when just thinking rather subjectively about such a situation one caneasily come up with between a few dozen and a few hundred possible controlledperceptions nearly all of which can influence the act of driving a car.
I can't help but think that when driving a car (as with many othersituations), there are possible perceptions associated with continuedexistence, lack of injury or pain, etc. that normally are quite well withintheir respective control references. Coming around a curve at 60 MPH and seeinga truck across the road combined with the perceptions that 1) the car will notstop before reaching the location of the stopped truck, 2) I have seen manymangled cars, 3) I have seen accident victims, 4) I have seen and known whatWERE accident victims but now they are classified as dead, etc., is likely toprovide a disturbance to the perceptions that, among other things include, theidea that I will enjoy the rest of the day.
I suspect that an error signal caused by a perception that maybe I am aboutto die, is likely to cause related control systems to engage in most vigorouscontrol activity. I do not think that "learned behavior" (experience)necessarily "goes out the window" but many people have had little experience indealing with life threatening situations. In such cases, their "behavior"probably should be similar to the infant.
-bill
Date: Sat Jul 16, 1994 2:37 pm PST
Subject: Disturbances
[From Bill Powers (940717.0620 MDT)] Paul George (940715.1145)
Your description of the constraints under which process control engineerswork makes the whole thing sound either (a) challenging or (b) infuriating.It's probably both.
On disturbances:
I'm very puzzled about your remarks concerning disturbances of controlledvariables. We don't seem to be talking about quite the same thing. Isaid
>> So am I to take it that in your process-control models, you do NOTmake an explicit provision for unpredictable disturbances to affect thecontrolled variable?
And you replied,
> Yup. We are forced to try to anticipate all possible problems anddeal with them. In fact for safety critical systems we are legally requiredto. That is why requirements analysis and design are so difficult. Ifsomething unanticipated does occur and something bad happens, we can be heldlegally accountable if we 'knew or should have known' that it couldoccur.
By "disturbance" I don't refer to "problems," but just normal effects thatthe environment would have on controlled variables if there were no control. Inour Figure, the disturbance is shown explicitly to allow for such effects. I'mhesitating here, because if you already know what I'm about to say, this wholecomment will seem silly and patronizing, but if you don't it's very importantto explain. I can't tell from your words which is the case.
Suppose you're controlling temperature (qi) of a sample of something. Thisis done by sensing (Fi) the temperature to produce a temperature signal (p),comparing (C) that signal with a reference-signal (r), and converting the error(e) via an amplifier (Fo) to send power output (qo) to the heater coil. Theoutput of the heater coil affects the sample according to the sample's heatcapacity and mass (Ff includes the heater coil and thermal constants of thesample). Also affecting the temperature of the sample is the temperature of thesurroundings (d), which drains heat from the sample according to theconductivity, radiation properties, and convective properties (Fd) of theintervening medium; this affects the sample temperature according to the samethermal properties of the sample.
So we have identified all the variables, signals, and functions in ourstandard Figure.
OK, the control system will vary qo (power to the heater) until the sensedtemperature matches the reference temperature. At this point there is anactively-maintained balance between the heat input from qo and heat losses tothe sample's surroundings. This is a continuous control system, not a bang-banghome thermostat.
Note that we can't even make this system work without the disturbance. Thedisturbance in this case consists of heat losses to the surroundings of thesample. Without those, the control system couldn't cool the sample when it istoo hot; it can only heat the sample.
The disturbance d is, in general, a variable; the temperature of thesample's surroundings can vary unpredictably. But we don't need to predictthose variations or even know what is causing them, because we have atemperature control system that acts directly on the sample temperature. If thesurroundings warm up, the resulting error will immediately reduce the heateroutput, preventing the temperature from changing more than a minuteamount.
What you call a "problem" would be, for example, a situation in which thesurroundings came to a temperature higher than the set-point of the temperaturecontroller. The control system is unable to handle this situation. If that isat all likely to happen, the control system would have to be redesigned with anoutput function capable of heating or cooling the sample according to the signof the error signal. Then the only "problem" that could occur would be for theambient temperature to become so extreme that the maximum heating or coolingcapacity was exceeded.
A "problem," as I'm defining it here, is a situation that goes outside therange of control of the control system. All closed-loop control systems arelimited as to the amount of output they can produce and the speed with whichthey can vary the output between maximum and minimum. That defines the universeof disturbances that the control system can handle automatically, simply out ofits basic design. Any disturbance that goes outside that space-time envelopeconstitutes a "problem," because control will fail as long as the disturbanceremains outside the envelope.
A "disturbance" is not a problem if it does not call for more output orfaster changes in output than the control system can produce. When I speak ofdisturbances, I'm talking about normal operation of the system. That's how Idefine a "normal" disturbance: one that does not exceed the capacities of thecontrol system.
It's not necessary to understand what is causing normal disturbances.Neither we nor the control system needs to know that. If the comparatorgenerates an error signal for ANY reason, whether an external disturbance or achange in the reference signal, the control loop will immediately correct it bymatching the perceptual signal to the reference signal (language requires us todescribe this process sequentially; it is really a simultaneous rebalancing ofall variables in the loop).
So: where do we stand on the subject of disturbances now?
Best, Bill P.
Date: Sun Jul 17, 1994 9:20 am PST
Subject: Re: Control of output? No such thing
<[Bill Leach 940717.11:27 EST(EDT)] >[Rick Marken(940715.1530)]
> ... All the control loop needs is a perceptual representation of thevariable under control -- the position of the door in this case -- which isvariable F in Bill's diagram; it is called the controlled variable inPCT.
Rick; I think that this statement could lead to some confusion. What youare saying is (I believe absolutely correct) but it is predicated upon thesimple idea that "position of door" is all that is being controlled and Isuspect such is rarely the case.
For various different reasons, we likely DO perceive the force that weapply to operate a door and have an active control loop for just thatperception. It is probable that the reference is a moderately loose reference(that is very low gain until the applied force rises to near some referencelimit).
I suspect that the reference is set based upon experience with similar (orprevious, if the same door) attempts to open the door.
This whole thing can get so complex when used as an example even though theexample appears so simple. For example, when one pushes upon a door to have theperception for the door match the reference (in this case door open) and thedoor does not move there are literally a multitude of perceptions that might beaffected.
One perception is that of "balance" as a result that the applied force willas per Newton's laws be equal in both directions and since the door did notmove...
Now then, while I might appear to be "challenging" what you are saying, Idon't believe that I really am. This "knowing the force applied" has nothing todo with the actual perception "door open" that is desired (except that thelower level control loops associated with actually operating the muscles dealwith such matters of course) but rather it is associated with a whole host ofother perceptions that are also under control. This would include even suchperceptions that one could break something if too much force is applied (saylike opening the plastic door on a piece of equipment).
OTOH, this does appear to me to be an excellent discussion of the conceptof the TEST. A control system DOES NOT need to know the force exerted tocontrol the perception "door open". If one observes someone "opening a door";1- the door does not open and 2- the person does not exert the maximum physicalforce that they can to achieve the an "open door condition" then (as Iunderstand it) the TEST tells us that there are other perceptions under controlhere that are related to the one that we are testing.
-bill
Date: Sun Jul 17, 1994 11:10 am PST
Subject: Mary on B:CP
from Mary Powers 940717 To Bill Leach:
Despite Tom Bourbon's statement, that behavior is the means by whichperceptions are controlled, you still, 940715, want to maintain your positionthat "behavior results from the control of perception".
I hope this is a semantic rather than a conceptual difference. But from thePCT point of view, once again, behavior is not a result, it is a means.Behavior results _in_ perceptual control, and is not a consequence ofit.
I can appreciate your desire to take as a starting point something that isa little easier to understand for those who are new to PCT. But I believe yourformulation is way, way down a slippery slope. It is the formulation thatCarver & Scheier and other "self-regulation" psychologists arepromulgating, and it misses the point entirely.
If behavior results from anything, it is from the discrepancy betweenperceived and desired/intended states. That is the part of the loop immediatelypreceding action. However, even this is probably an excessively lineal andsequential way of looking at a process which is essentially simultaneous allaround the loop.
There really are no baby steps to take between behavior as outcome,consequence or result, and behavior as the control of perception. One of thebig difficulties PCT has in making its way in the world is that you can't shapethe understanding of it incrementally. Either you force PCT data to fit yourworld view, or you make the jump, and the world never looks quite the sameagain. Nothing in between.
Mary P.
Date: Sun Jul 17, 1994 12:26 pm PST
Subject: Re: Mary on B:CP
<[Bill Leach 940717.15:42 EST(EDT)] >Mary Powers 940717
> There really are no baby steps to take between behavior as outcome,consequence or result, and behavior as the control of perception.
I am beginning to believe the truth of that statement. It seems to me thatmost of the differences in opinion between you, Bill, Rick, Tom and virtuallyanyone else (where such difference occur of course) rest squarely upon adisagreement with the truth of the following:
Behavior as a function of some environmental condition can not exist at allunless there is both a perception for that environmental condition and areference FOR THAT PERCEPTION. This is likely also true for "unobservable"behavior such as thinking.
No change in behavior can occur unless either the perception changes or thereference changes.
Now there are other matters such as reorganization are maybe even physicaldamage to the control system but such matters are usually mentioned in any postwhere applicable.
I admit to being a bit "dense" now and then or should that be... all thetime? :-) . . . I realize that you did not raise this question but only myassertion that "behavior results from the control of perception" and I guessthat I am trying to say that I am beginning to see were such "minor"differences are not so minor after all... they are rather the root ofmisunderstanding.
-bill
Date: Mon Jul 18, 1994 10:21 am PST
Subject: Re: Disturbances
[Paul George 940718 0930] [From Bill Powers (940717.0620 MDT)]
> So: where do we stand on the subject of disturbances now?
Just fine. I share your distinction between 'problem' and'disturbance'.
The engineer is usually interested in 1) keeping some controlled variablewithin a range (normal operation), 2) designing the system so that it has thenecessary sensors and instrumentality to keep the process under control, and 3)anticipating and detecting any circumstance that indicate the process isgetting out of control (i.e. a defect in 2). The latter is why control systemshave 'alarms' to detect potentially dangerous situations in areas not underadaptive control. 'Alerting' (ducking to avoid bricks from Rick Marken) is abig part of industrial control systems, of course they are alerting the humanpart of the control system. The operator may then take mitigating action byopening a valve, shutting off power, calling for evacuating a tri state area:-}, etc.
Date: Mon Jul 18, 1994 10:22 am PST
Subject: Re: Heretics, etc
[Paul George 940718 10:00] >[Rick Marken (940715.1300)]
> What's your guess at the correlation between number of neurons &connections and 'levels of intelligence'?
It's not my field of expertise, so my opinion isn't worth very much.However, complexity of behavior in species appears to be something of a stepfunction (possible multiple curves or sawteeth rather than right angel steps)associated with brain size. A certain critical mass seems to be needed tosupport a given level of complexity.
Date: Tue Jul 19, 1994 1:40 am PST
Subject: Re: Diagram Terms o
From Tom Bourbon [940718.1739]
Back from the wedding weekend with my credit cards melted down from overuse-- the true role of the father of the bride!
What a flood of mail on the net!
>[Paul George 940715 12:00]
>>[Bill Leach 940714.20:34 EST(EDT)]
>> Then again... there might actually be two types of situations. Iam thinking of the person that appears to really have two modes of operation ofthe brakes; Variable pressure control to slow the vehicle at some controlledrate is one and the other is "panic" mode where they just "press for all thatthey are worth" (hopefully at least on the brake but often not!).
> That is the situation I had in mind. People often react verydifferently to gaining on another vehicle, and seeing a truck stopped in frontof them.
Why would those two scenarios call for two different "modes" of control? Ifperceived rate of closure is being controlled, with a reference value of zero,or something very small, then think of how the error signals would be changingin the two scenarios, prior to the person applying the brakes. Even if theperson's gain did not change, or the "mode" of control remained the same, theerror signal would be much bigger and growing more rapidly in the one scenariothan in the other. Perhaps?
Later, Tom
Date: Tue Jul 19, 1994 1:51 am PST
Subject: Disturbing alerts
[From Rick Marken (940718.1430)] Paul George (940718 0930)
> The engineer is usually interested in 1) keeping some controlledvariable within a range (normal operation), 2) designing the system so that ithas the necessary sensors and instrumentality to keep the process undercontrol, and 3) anticipating and detecting any circumstance that indicate theprocess is getting out of control (i.e. a defect in 2).
I don't understand point 3). Are you trying to anticipate deviations of acontrolled variable from its reference state or are you trying to anticipatedisturbances to the controlled variable?
> 'Alerting' (ducking to avoid bricks from Rick Marken) is a big partof industrial control systems
Satellite control systems, too. Of course, a signal is only "alerting" ifthe operator has learned to treat the signal that way; "alertingness" isdetermined by the operator, not the signal.
Rick
Date: Wed Jul 20, 1994 1:05 am PST
Subject: Re: Disturbing alerts
[Paul George 940719 11:30] >[Rick Marken (940718.1430)]
>> 3) anticipating and detecting any circumstance that indicate theprocess is getting out of control (i.e. a defect in 2).
> I don't understand point 3). Are you trying to anticipate deviationsof a controlled variable from its reference state or are you trying toanticipate disturbances to the controlled variable?
I didn't phrase it very clearly. The system normally perceives more than itcan affect. In design I must anticipate different kinds of disturbances thatthe system must detect in order to control the process (sorry;-). Sometimesthat 'controlled variable' can be controlled (i.e. affected by an action of acontrol system) sometimes it can't. I must anticipate accident, ignorance, andmalice. Ideally I will design the control system so that the process _cannot_run away and no hazardous situation can occur. Otherwise I must detect thecondition and inform the operator, who (hopefully) can do something to correctthe situation, or prevent disaster, or at least mitigate the effects.
Kodan lieutenant: "We are trapped in the moon's gravitational pull, whatdo we do!!"
Kodan commander: "We Die."
{From the movie 'The Last Starfighter'}
Paul
PS: I finally got hold of B:CP, YEA!!
Date: Wed Jul 20, 1994 1:34 am PST
Subject: Re: Disturbing alerts
<[Bill Leach 940719.19:12 EST(EDT)] >[Rick Marken(940718.1430)]
>>Paul George (940718 0930) --
>> 3) anticipating and detecting any circumstance that indicate theprocess is getting out of control (i.e. a defect in 2).
>Rick:
> I don't understand point 3). Are you trying to anticipate deviationsof a controlled variable from its reference state or are you trying toanticipate disturbances to the controlled variable?
Among other things, in engineered control systems, there are oftenenvironmental conditions whose status is quite important to the control processbut is not controlled by the engineered system hardware.
An example that I can think of right off the top of my head would be abatch process tank that is manually made up by an operator. An engineeredcontrol system would often be designed to at least estimate when a new batch ofchemicals might be needed.
A different sort of example is the situation where a sensor has a lifetimeand it is difficult to determine the sensor's status. The control system thenmay attempt to create "lifetime statistics" for the sensor and even attempt toderive failure clues from experience with the actual operating process.
I would say that living control systems do the same sort of thing but 1) weare often probably not aware of it even when we are doing it ourselves and 2)we are so damned adaptive at times that even a careful observer might notnotice a change that we make to continue to control a perception when something"goes wrong" with whatever method that we were using.
-bill
Date: Wed Jul 20, 1994 1:36 am PST
Subject: Cause systems
[From Rick Marken (940719.1400)] Paul George (940719 11:30) --
> In design I must anticipate different kinds of disturbances that thesystem must detect in order to control the process (sorry;-).
I think you're going to have to diagram one of these controllers of yours.In a real control system, there is no need to detect or anticipatedisturbances. I thought that you agreed with Bill's definition ofdisturbances?
> Ideally I will design the control system so that the process _cannot_run away and no hazardous situation can occur.
I think we need a diagram of what you call a control system. Could you giveus a diagram of one of the "control systems" you build at work, identifyingcontrolled variables, disturbances, etc.
I'm getting the distinct impression that what you call a "control system"is not a control system . A lot of people you the word "control" as a synonymfor "cause"; could this be what's going on in "process control"?
Best Rick
Date: Wed Jul 20, 1994 1:38 am PST
Subject: Re: Gestalt
From Tom Bourbon [940715.1722]
I'm way behind with my replies, but here goes a start at trying to catchup.
>[Paul George 940715 10:30] >>Tom Bourbon [940714.1750]
>Paul:
>>> Perhaps you should give me your definition of'gestalt'.
Tom:
>> Ah, it was _you_ who introduced the term into the discussion. I'vecalled _your_ hand. Show your cards! :-)
Paul:
> I've been trying. And from your discussion I think we are saying muchthe same thing. One more time.
> To perceive something which exists in the environment, one needs aset of inputs. The blind men couldn't correctly perceive the elephant becausethey didn't touch enough of it and couldn't see the whole. Similarly, I can'tclap with only one hand. It isn't the number of components (def differsbetween examples) that matters, it's just that if you don't have the completeset, nothing happens. At some point you have enough knowledge, or controlloops, or perception, or instrumentality to affect the environment in a waythat allows you to control your perceptions effectively (hope that was badenough wording :-). Adding more may increase your level (effectiveness) ofcontrol, or may not. But less means you can't control at all. At some pointthere is a critical mass.
I Believe I'm beginning to see why you thought you were telling me what youmean by a "gestalt" and I didn't realize that was the case. I think of"gestalt" in terms of the historic Gestalt "school" in psychology. Itsadherents urged that perceptual experience is _always_ "whole" and perceptionis different from elementary sensations -- an idea that in some inexplicableway became the textbook chestnut, "the whole is greater than the sum of theparts."
They also made a big thing of the idea that perception and neurologicalevents are "isomorphic" and that both neural events and perception are entirely"contemporaneous:" they are simultaneous and in the immediate present --nothing from the past or future can enter into present neurological andperceptual events. (I guess that would rule out "feedforward," wouldn't it?;-)
In light of that tradition, the story of the blind men and the elephant isnot a story about a failure of three people to have their individual"gestalts." On the traditional Gestalt reading, each of the blind men had theonly perceptions he could have: those that were isomorphic with theorganization and functional state of his nervous system, _whatever_ those mighthappen to be. Each blind man had "whole" perceptions -- "gestalts," if you like-- not incomplete ones. Only a differently organized person (say one who wasnot blind) could have whole perceptions that were different from those of eachof the blind men. And who is to say the sighted person sees all there "reallyis?" Different organizations, different activity, different perceptions -- andall of them "whole."
> A Baby can't act very successfully because it hasn't learned enoughabout what the sensory stimuli 'means'.
And the organization isn't there to support "mature" actions; but babiesfunction as whole systems, with whole perceptions and actions. (I like thismore accurate reading of Gestalt theory -- it flies in the face of morefamiliar and popular ideas that infants are "incomplete" adults and that theymust "develop" into adult "finished products.")
> I can't form complete sentences until I know enough words andgrammar.
But you do whole things other than form complete sentences . . . And soon.
> . . . In general there is some correlation between the number ofneurons & connections and 'levels of intelligence' (speaking of fuzzyterms).
Perhaps, in a very loose sense. That puts whales far ahead of us, I guess,and . . . Hmm. Maybe I won't even stand on my "perhaps." Whatever the case,the Gestaltists would have come back with the idea that no species is a partialrealization of any other species; each is complete and its experiences andactions are "whole" for its particular organization.
> Hope these examples clarify what I am talking about.
I think they did. I hope my reply clarifies why your use of "gestalt"didn't look familiar to me. ;-)
Later, Tom
Date: Wed Jul 20, 1994 1:44 am PST
Subject: Modes of Control
[Paul George 940719 10:40] >Tom Bourbon [940718.1739]
> Why would those two scenarios call for two different "modes" ofcontrol? If perceived rate of closure is being controlled, with a referencevalue of zero, or something very small, then think of how the error signalswould be changing in the two scenarios, prior to the person applying thebrakes. Even if the person's gain did not change, or the "mode" of controlremained the same, the error signal would be much bigger and growing morerapidly in the one scenario than in the other. Perhaps?
Perhaps. Sometimes the magnitude or rate of change of the error signal maybe sufficient. I guess it depends how sophisticated you allow the outputfunction (F2) to be. If it can select between different actions based uponthresholds for the error signal, fine. The idea is that different actions maybe required in response to different amounts or rates of disturbance. To dealwith a punch, sometimes you block, sometimes you duck. Depends on how fast itis coming, and where your arms are.
I think the more common structure might be having another control loopmonitoring the error signal, treating it as a controlled variable. That loop'soutput function would then activate the appropriate loop for the needed kind ofaction (i.e an 'emergency' loop). This presupposes that you have multiple loopscapable of controlling the same perception, but have their reference variablesadjusted so only one actually generates an error signal at a given time.Possible?
Paul
Date: Wed Jul 20, 1994 5:12 pm PST
Subject: Process control systems
[Paul George 940720 11:00 ] >[Rick Marken (940719.1400)]
Me:
>> In design I must anticipate different kinds of disturbances thatthe system must detect in order to control the process (sorry;-).
>> Ideally I will design the control system so that the process_cannot_ run away and no hazardous situation can occur.
> I'm getting the distinct impression that what you call a "controlsystem" is not a control system . A lot of people you the word "control" as asynonym for "cause"; could this be what's going on in "processcontrol"?
I think this is just mis-communication. Having worked with satellitecontrol systems at Loral Aerospace, they seem much the same as process control- from the control system's standpoint.
The concept is that when we _design_ a control system we must determinewhat needs to be controlled and what can go wrong. For a satellite, if weconsider a loss of carrier or orientation to be significant, we had better havesensors to detect it. Further, it had best have some mechanism for stationkeeping and re-orientation. Then, we has better consider how the attitudecontrol jets or controls could screw up (say lock on, vent the hydrazene,overheat fuel,.....) and design in preventive measures or sensors to detect theevent.
Paul
Subject: Re: Disturbance Control, Political Steering
Paul George 940720 17:30
>[From Rick Marken (940720.0845)] >>Bill Leach (940719.19:12)--
>> Among other things, in engineered control systems, there are oftenenvironmental conditions whose status is quite important to the control processbut is not controlled by the engineered system hardware.
> I'd say that that's true of ALL control systems; the environmentalconditions of which you speak are called disturbances; they are the reason forbuilding a control system in the first place -- to protect variables from theunpredictable effects of these conditions.
They are called disturbances _after_ the system is constructed. Duringanalysis or design there are no controlled variables _yet_. That is the purposeof the design - to determine what input functions, output functions, andhierarchical levels of control are needed to meet the purpose of keeping theprocess stable and predictable.
> This estimate is not really part of the control process; it is justanother variable in the environment
Here is the major terminology conflict. To us 'control system' is more thanthe controller. It includes the equipment, sensors, and multiple controllers.(so is a 'satellite control system') It even sometimes includes the operatorvia interaction with displays as well as instruments or controls not hooked tocontroller input or output functions. In biology one can distinguish betweenthe nervous system and everything else. You could also draw a system boundaryat the skin. A process control system is more like a work party than aperson.
Paul
Date: Wed Jul 20, 1994 7:24 pm PST
Subject: Comments to Paul George (from Mary)
[from Mary Powers 940720 8:45] Paul George:
> Sometimes the magnitude or rate of change of the error signal may besufficient. i guess it depends on how sophisticated you allow the outputfunction (F2) to be. if it can select between different actions based uponthresholds for the error signal, fine. the idea is that different actions maybe required in response to different amounts or rates of disturbance. to dealwith a punch, sometimes you block, sometimes you duck. Depends on how fast itis coming, and where your arms are. I think the more common structure might behaving another control loop monitoring the error signal, treating it as acontrolled variable. that loop's output function would then activate theappropriate loop for the needed kind of action (i.e. an 'emergency' loop. thispresupposes that you have multiple loops capable of controlling the sameperception, but have their reference variables adjusted so only one actuallygenerates an error signal at a given time. Possible?
Possible, maybe, if you are designing such a system. But remember PCT isunder the constraint of being neurologically plausible, and also under theconstraint of keeping things as simple as possible unless evidence demandsotherwise. Therefore the output is not allowed to be very sophisticated.
Part of the problem here is that you are looking at the canonical, singleloop model, and trying to see how it can do the work of what PCT conceives ofas a hierarchical arrangement of a multitude of control systems, none of whichis engaged in monitoring the size of the error signal (aside from thepostulated reorganization function which has as input a state of large andchronic error).
In PCT, error signals are outputs to lower levels, where they function asreference signals. Only at the lowest level is output a signal for muscles tocontract. There is no function selecting actions: the actions are dependent onnumerous output signals to various muscles, which in turn are dependent on thecomparison of the desired state of affairs to the perceived environmentalsituation. In the case of ducking or blocking a punch, the state of theenvironment is the perception of how fast the punch is coming, PLUS, as youcorrectly pointed out, where your arms are - which is as much a part of theenvironment of the system as the approaching punch. If your arms are in aparticular place, you've got to duck because it's too late to block,etc.
In the heat of a fight, this is all going on at pretty low levels, and yoursuccess at blocking or ducking is not a high- level intellectual exercise -it's a matter of much practice and training to have lower-level loops -perceptions, comparison, outputs - be sufficient and correct; for the lowersystems to duck rather than block because of experience with what can be doneor not done when the arms are in various positions. The same applies to dealingwith braking in emergencies. Hitting the brakes hard, which in normalsituations is the right thing to do, is a poor idea on an icy slope, which iswhy incomers from Texas and California end up in ditches around here in thewintertime more often than the natives.
Mary P.
Date: Wed Jul 20, 1994 9:56 pm PST
Subject: B:CP impressions - 1st cut
[Paul George 940720 17:00]
Having read through the first 8 chapters, I begin to see the source of themisunderstanding of my posts, as most discussion seem to focus on the firstthree levels (orders) of the CNS. PCT does agree with my mental model, therewere just some scope and terminology issues. All in all a great piece ofwork.
I guess it's lack of acceptance is due to perceptual filtering, or morelikely that 'other schools' never actually _read_ (with understanding) thebook. I found Bill Power's post on the e. coli paper review committee revealingas to the nature of 'peer review'. However you should be careful that you don'tfollow the same pattern in judging other's work. It is important to evaluate apiece of work using the _author's_ frame of reference, not your own.Translation is your job when you are not using terminology in a 'standard' wayor when using 'non-standard' concepts.
May I suggest that the 'FAQ'/intro add a summary of the nature of neuralcurrents, the types of neural circuits, and the distinction between first orderand higher order systems. Since B:CP is relatively hard to find (lacking moneyto order it from the Powers), this would avoid a lot of confusion. {I had toget my copy via inter-library loan from a Cincinnati public library (I live inCleveland). It apparently wasn't available from the universities in the area,of which there are many)
The essence of PCT is the functioning of organisms (a normal focus ofpsychology as opposed to sociology). It can be summarized thusly: All actionand sensation is produced via the interaction of neurons and muscles. Further,the central nervous system is composed of neurons. It therefor follows that anyperception or behavior, regardless of complexity, must be producible via theinteraction of neurons through neural currents (in the absence of some othermechanism - 'a ghost in the machine'). We can demonstrate that negativefeedback control is the mechanism used for first order interaction with theenvironment, and at the level of spinal reflex. We can use models to show itcan work for higher order behavior. Thus, by default we presume it is themechanism used at all levels.
When I talk about a system or control system I am usually looking at morethan one entity. A biological analogy to a process control system would be asupervisor monitoring 2 operators, who are each operating a set of toolsinvolved in a manufacturing process. This is similar to HPCT, except thatneural currents are not the _only_ mechanism for interaction of the 'nodes'. Idon't think that this really changes anything, except by adding 1st level inputand output function errors and adding the possibility of non pulse codedsignals.
Please bear with me on the following as I am using a 20 year old work.There may have been elaborations in the mean time, but I presume there would bea '2nd edition' if there were major changes.
An observation: error signals don't theoretically have to only havepositive values from an information standpoint. You can cheat by 'biasing' thesignal. If the signal may vary from 0-30 ppm, I can set the comparator andoutput function so that 15 = no error (a logical 0). I could then set an 'upperthreshold' at 25 and a lower threshold at 5. The output function could use thisinformation to 'select' the proper action or magnitude of action. This couldallow a stepped sawtooth output function rather than a continuous one. I'm notsaying that this _ever_ happens in biology, or that it can't be implementedwith a network of simple nodes, just that it would work. This might be usefulfor designing automata using PCT.
Question: why is it illegal to pass an error signal (directly or through a'repeater' output function) to another node as a 'sensation'? While the analogymay not bear close examination, pain might be something of this type. Theoutput is just passed up to tell a higher level controller that something iswrong, and how badly. It can be used to indicate that 'control' is not workingand another strategy must be applied. It is up to the higher level to setreference levels elsewhere in the network so that the error signal ismitigated. It's not very useful at the lowest couple of orders, but might beuseful at the 'cognitive' levels of control. This also might correspond to'alerting'. It is just another input signal, but generated by a comparatorrather than a lower order input function. Again, I don't assert that itactually exists in nervous systems, or that it is necessary. OTOH inengineering we sometimes find that a more complex structure is more efficientthan the equivalent constructed from simple structures. If nothing else youhave propagation and processing lags. Biology has a limitation in that itdeveloped by 'growing like topsy'. New levels and nodes were added to existingones that worked.
As your thinking has evolved over 20 years, do you have any problem withhaving 'neural logic' as a part of input, output, or comparator functions? Isthere any theoretical problem with a single node (or subsystem) being atdifferent levels of the hierarchy with respect to other nodes or subsystems{Higher apparently means 'sets another's reference level'} at the same time? Mymental model can envision a network rather than a true hierarchy
On another minor matter, I am not sure that nodes interact only throughneural currents in biological systems. Some output functions result in therelease of hormones, neurochemicals or substances like adrenaline (blanking onthe name). These are certainly sensed by other nodes, and not necessarilyhierarchically (i.e 4th level affecting 4th or other level). I guess this couldbe viewed as an 'environmental' interaction of physical laws, but seems to memore like a 2nd or higher order interaction. I'm not sure it really matters ifa signal is transmitted in terms of neural current frequency or in terms of achemical concentration which must be translated via an input or referencesignal transducer function. Again, it may be just where you draw the systemboundary.
Re Tom Bourbon [940719.1202]
I don't think creating a taxomy of types of signals or types of controlnodes is silly in and of itself, any more than distinguishing between orders ofcontrol systems or sections of the brain. Yea it is all 'just control' or 'justneural currents', but the distinctions are sometimes useful. We wouldn't getvery far in medicine or physiology if we fixated on the idea that 'cells arejust cells' and ignored their differentiations and groupings (e.g. organs). Wemay be able (and have) to define standard structures or patterns used forvarious 'types' of perception and control. The question is which groupings ordistinctions make sense.
Enough digital diarrhea for today, Paul
Date: Wed Jul 20, 1994 10:02 pm PST
Subject: Re: B:CP impressions - 1st cut
<[Bill Leach 940720.23:22 EST(EDT)] >[Paul George 94072017:00]
> Having read through the first 8 chapters, I begin to see the sourceof the misunderstanding of my posts, as most discussion seem to focus on thefirst three levels (orders) of the CNS.PCT does agree with my mental model,there were just some scope and terminology issues. All in all a great piece ofwork.
Delighted to see that you have started into B:CP. Your difficulty inobtaining it is certainly a sign of at least part of the problem.
> I guess it's lack of acceptance is due to perceptual filtering, ormore likely that 'other schools' never actually _read_ (with understanding)the book.
I am about willing to even make a bet, that few if any in a position ofauthority have read the book at all much less study the concepts.
> It is important to evaluate a piece of work using the _author's_frame of reference, not your own.
This is hardly correct. It is essential to attempt to understand theauthor's meaning(s) but at some point you must evaluate the author's work basedupon some standards of your own. The idea that a technical book is acceptedbased upon the author's remaining consistent his own "frame of reference" isabsurd.
In the first place, the author may state goals for his work that are notmet even though he maintains steadfastly that they were. The stated goals maynot be the true purpose of the work. And the work may rely upon principles ortruths that are not. All of this must be considered when doing a scholarlyreading.
Admittedly, it is easy to "rationalize" and not do an honest evaluation ofwhat the author actually has said. I believe in the case of PCT the matter isreally much simpler in that so few people actually understand the phenomenon ofcontrol that they quickly reveal errors when discussing control.
> Translation is your job when you are not using terminology in a'standard' way or when using 'non-standard' concepts.
This is true and it is also true that it is the "job" of anyone trying tounderstand the work of another to attempt to come to terms. A particulardifficulty in this area for PCT is that almost NO ONE ELSE uses precisemeanings for relevant terms. In the behavioral sciences the statement is truealmost without exception but unfortunately even in the engineered controlsystems field terms are rather loose.
> The essence of PCT is the functioning of organisms (a normal focus ofpsychology as opposed to sociology).
Well, if sociologists had some concept that they are dealing with livingcontrol system and the implication of such a condition, they would have a greatdeal of interest in the matter.
> It can be summarized thusly: All action and sensation is produced viathe interaction of neurons ...
... Thus, by default we presume it is the mechanism used at alllevels.
I think "them's fighten' words" :-) PCT is about the idea that Behavior isthe control of perception. It does not matter that it actually appears thatneural comparators do not allow sign switching, it does not matter ifreferences and their perceptual signals might be biased or might not. PCT doesis not only not "bothered" by the idea that many signals may be chemical innature (as opposed to electrical) but the theory even helps explain thefundamental operation of some of them and further just about requires theexistence of some sort of chemical type system for reorganization to occur inthe fashion that the theory predicts. The real point is that whatever it is,there is overwhelmingly strong evidence that it is a closed loop negativefeedback system and that is what PCT is about.
> When I talk about a system or control system I am usually looking atmore than one entity.
PCT does not have a problem in dealing with multiple controlsystems.
> Question: why is it illegal to pass an error signal (directly orthrough a 'repeater' output function) to another node as a 'sensation'?
I would suggest that it is probably NOT illegal in any sense of the wordbut rather it does not happen because it makes no sense from a control systemstandpoint to do so. When controlling, the error is so close to zero to beuseless for any other purposes (it is in the noise). The perception itselfcould also be an input to another perceptual control function that could handleproblems associated with loss of control by the first system. Additionally,there does appear to be a system for sensing loss of control.
> While the analogy may not bear close examination, pain might besomething of this type.
If I am not misunderstanding you here, Pain is clearly not a good example.Pain happens to be one of the areas where a great deal of medical research hasbeen conducted. Though certainly not all pain has been studied this way butmany studies indicate that mechanics of pain is the result of sensors sendingsignals up through the nervous system to the brain. There is no indication thatthere is any connection between these signals and the output of any othersystem.
> OTOH in engineering we sometimes find that a more complex structureis more efficient than the equivalent constructed from simplestructures.
I should like an example of this. This sounds a bit like the argument thatit is impossible to write a program in assembler that is as efficient andeffective as one written using an optimizing high level compiler. Such astatement is of course pure bunk. Attempting such a thing would be inefficient,would require one hell of a programmer, be impossible to maintain, etc. BUT ithas to be doable since the optimizing compiler is using the very same "simplestructures" that are available to the assembler.
> If nothing else you have propagation and processing lags.
What do you mean by this? Higher level engineered control structures don'thave propagation and processing lags?
I would like to suggest that in using the high level development tools thatyou have at your disposal for designing control systems, what is happening isthat 1) you have a great deal more knowledge available to you about thecapability of the elements of the system and you have the ability to easilycontrol perceptions that you would not likely attempt to control if you weretrying to use only simpler controllers for the job.
> On another minor matter, I am not sure that nodes interact onlythrough neural currents in biological systems. Some output functions result inthe release of hormones, neurochemicals or substances ...
Keep reading :-)
-bill
Date: Thu Jul 21, 1994 11:29 am PST
Subject: Re: Comments to Paul George (from Mary)
[Paul George 940721 10:00] [Mary Powers 940720 8:45]
> In PCT, error signals are outputs to lower levels, where theyfunction as reference signals. Only at the lowest level is output a signal formuscles to contract. There is no function selecting actions: the actions aredependent on numerous output signals to various muscles, which in turn aredependent on the comparison of the desired state of affairs to the perceivedenvironmental situation.
Sorry, sloppy use of the term 'action'. What I meant was a higher levelloop using 'neural logic' or some kind of rule set to select _which_ of a setof lower loops should have its reference levels adjusted to what degree,'causing' them to control (generate outputs). Ultimately action occurs whichhopefully causes the original error signal to diminish.
See also my post yesterday on B:CP
Thanks for the response Paul George
Date: Thu Jul 21, 1994 11:38 am PST
Subject: Misc comments
[From Bill Powers (940721.0815)] Paul George (940720.1100)
You (and Bill Leach) are making some good points about the differencesbetween process control designs and modeling people. A lot of what goes on inthe design of artificial control systems doesn't translate into PCT. The realaffinity between the approaches is at a deeper level; the phenomena ofclosed-loop systems, basic ideas like perceptual representation, comparison,action, external feedback paths. One difference is that the process controlengineer knows too much; he/she knows what is causing disturbances, what mightcause disturbances, and what sorts of countermeasures can be taken before thedisturbances ever happen. To an organism, everything comes as a surprise (otherthan the very few behavioral systems that are in something like workingcondition when we're born).
I'm delighted that you're making your way through BCP. It is hard to get;distributors don't stock it. Aldine-deGruyter has done a new printing (don'tknow how large), so it's available from them for, I fear, about $43. Totalsales so far must be somewhere between 5000 and 7000 copies.
I haven't done a revision, though it could use one. Just haven't got theinspiration to try. The internet absorbs all my creative writing thesedays.
> I found Bill Power's post on the e. coli paper review committeerevealing as to the nature of 'peer review'. However you should be carefulthat you don't follow the same pattern in judging other's work.
Right you are. I always make a pretty serious attempt to grasp what theother person's model is before saying anything about it. But you can go only sofar in following another person's reasoning, especially when the person doesn'treally understand what's required of a model.
The biggest problem I have usually comes right at the beginning, where thespeaker wants us to accept certain vital premises for the sake of the rest ofthe argument. If I have no knowledge or opinions about the premises, I'll playalong, but it gets really difficult for me when a premise is something I justflat can't believe. Sometimes the premise is right there in the title: "Controlof responses by discriminative stimuli as a function of frequency ofreinforcement." How am I going to get past that? I just can't. I have the sameproblem with complex mathematical developments. I find them difficult enough,but when they start out by assuming things about which I am very dubious I justcan't get up the energy to play out the game, even assuming that thederivations and lemmas and theorems are free of mistakes. This has irritatedsome people; they say "Well can't you just grant that for the sake of theargument, and see what we can derive from it?" I say, "I'll grant it when yougive me some reason to," and there goes a beautiful relationship down thedrain.
That's one reason I'm so big on demos and experiments. That's how youestablish that your premises are reasonable. Psychological theorists don't seemto pay much attention to that sort of thing.
Glad you liked the section on neural currents and computations; not manypeople read that chapter. At least I can claim that I was thinking about analogneural nets before 1973.
> All action and sensation is produced via the interaction of neuronsand muscles.
OK, if you'll add " ... through effects of the muscles on theenvironment."
> ... is similar to HPCT, except that neural currents are not the_only_ mechanism for interaction of the 'nodes'.
A realistic model of nervous-system operation would have to include thebiochemistry of neurons. That is implicit in the notion of neural computingfunctions, but not spelled out in PCT. Too detailed a level of analysis for ourcurrent state of understanding.
We are commencing some work on purely biochemical control systems. Staytuned.
> An observation: error signals don't theoretically have to only havepositive values from an information standpoint. You can cheat by 'biasing' thesignal.
Right. The model we use is a "canonical" model, meaning that there are manyother forms that would be equivalent to it in function. What you're suggestingare variations that could well exist, but which would be equivalent, in theend, to the canonical model. In the brain stem, the reference signals appear toenter the output function along with the perceptual signals (at least in theRed Nucleus), and there is no distinct comparator. But the actual arrangementcan be reduced to the canonical model by redefining some functions andconstants.
> Question: why is it illegal to pass an error signal (directly orthrough a 'repeater' output function) to another node as a 'sensation'?
When you start talking about rerouting error signals to other systems, thenthe architecture can't be reduced to the canonical model. That may be perfectlyOK, but first you have to demonstrate how a system organized that way wouldactually work, and show that there's a behavioral phenomenon that needs it.There are lots of possible variations on the current model, but until they'vebeen simulated and applied as quantitative explanations of real behavior, theyhave to remain "unofficial." There are lots of unofficial ideas floatingaround, some from me, but they're just possibilities so far. Until we're forcedto explore them more seriously (because the existing model runs up against aphenomenon it can't explain and that we're willing to do the work on), theywill continue to float. There are plenty of simple phenomena that the canonicalmodel, alone or in hierarchies, handles very well. We're still at the pointwhere we're very happy to do simple experiments that work out. We don't acceptchanges to the model that aren't backed up by some pretty solidevidence.
As you have guessed, most of what we say about higher levels of control isin the realms of entertainment. We can't really model those levels oforganization yet.
As a system designer, you have to watch out for designing systems asopposed to modeling them. Any clever designer can think of six differentdesigns to accomplish a given result before breakfast, especially when workingwithout any factual constraints. What we're trying to do is figure out how theREAL system accomplishes what it does, and in that case there could be sixdesigns that would work and all of them are wrong. This is not something thatcontrol-system engineers normally have to worry about.
> As your thinking has evolved over 20 years, do you have any problemwith having 'neural logic' as a part of input, output, or comparatorfunctions?
No, and I never did. You'll be running across a "program" level of control,which I envision specifically as using symbolic computations, whether logicalor of any other kind (like calculus). I see the lower levels as exclusivelyanalog in nature, but if some phenomenon arises that calls for digital logic ina low-level system, there's nothing in principle against it. All anyone has todo is demonstrate that it's needed.
> Is there any theoretical problem with a single node (or subsystem)being at different levels of the hierarchy with respect to other nodes orsubsystems {Higher apparently means 'sets another's reference level'} at thesame time?
Yes, because I see the hierarchy as physical, not conceptual. The lower-level control systems are located physically near the periphery of the nervoussystem (the spinal and brain-stem reflexes, for example). The pathways thatconnect different levels are fairly well known up to a point, and they resemblethose in the hierarchical model (not by accident). There's nothing conceptuallyto prevent the output of, say, a first-level control system from becoming thereference input of a sixth- level system -- except that the outputs offirst-level systems actually go to muscles, not to higher systems. There areneuroanatomical constraints on the model. They aren't obvious in the finalproduct, but I did do quite a lot of study of that literature while putting themodel together, looking for constraints that would narrow the possibilities.There isn't as much useful information there as I had hoped for, but there issome.
> My mental model can envision a network rather than a truehierarchy.
Yes, that thought has nagged at me for a long time. I keep thinking ofcases where what I call a third-level control systems seems to operate via afifth-level control system, and so on. So far I've been able to resolve most ofthese possibilities, but the question is still open whether we should think ofthese "levels" as "dimensions" instead. This possibility was put to me by myfriend Kirk Sattley, a lurker on this net, in about 1954, on the very day whensomething else he said resulted in the picture of the control hierarchysuddenly falling into place. I opted for the hierarchy, but have wondered eversince whether there isn't something to the other view, too.
Whatever the case eventually proves to be, there will be constraints on thesolution. You can mess around with network ideas and come up with lots ofinteresting-looking designs, but "interesting" doesn't mean "right."
> Some output functions result in the release of hormones,neurochemicals or substances like adrenaline (blanking on the name).
There are many hormonal control systems at the level of organs, and belowthat many control systems of more detailed nature, all the way down to RNA andDNA. I see these as part of the environment in which the behavioral (neural)systems live. At the level of the hypothalamus, there are neural signals whichgo into the pituitary, where they seem to serve as reference signals for manyhormonal control loops (control of circulating thyroxin is one well-knowncontrol system of this kind). There are also many neural sensory signalsoriginating among these organ and biochemical systems. So to the neuralhierarchy, these systems are simply another part of the environment that can besensed and affected via neural signals. Of course we give different names tothe perceptions arising from inside the body; feelings and emotions and otherproprioceptive things. I think of these control systems as being below thefirst level of control in the neural hierarchy. Not much has been done withthem, by me, although others have studied isolated systems at thesephysiological levels.
> I'm not sure it really matters if a signal is transmitted in terms ofneural current frequency or in terms of a chemical concentration which must betranslated via an input or reference signal transducer function.
The neural currents interact chemically with neural cell bodies at thesynapses. Inside the cell bodies, the chemical influences of many simultaneousincoming neural signals, represented by neurotransmitters, interact accordingto laws of chemistry and diffusion to produce new signals at the output thatcan be complex functions of the input signals. A single neuron can be quite acomplex analog computer.
The literal _transmission_ of neural signals is an electrochemicalphenomenon, and the all-or-nothing nature of the impulse says that informationis most likely to be carried by frequency variations. The concentrations ofneurotransmitters at synaptic junctions, and even more so inside the receivingcell body, vary on a longer time scale than a single impulse, so the naturalrelationship between the chemical processes in the cell body and the neuralsignals is one of frequency- to-concentration conversion. The physicalchemistry in the nerve cell is really the temporal bottleneck that says anindividual impulse has no significance. There are others who disagree with me,claiming that neural signals can be multiplexed into a single fiber (anddistinguished from each other afterward!), and that information might betransmitted by small variations in amplitude of spike as well as frequency. Butthe simplest view is that neural signals are measured as continuous variablesin terms of frequency. Maybe this isn't true everywhere; there are all kinds ofsynaptic effects other than my basic model, although they are not common. Idon't really think we need to carry the model to that level just now.
I am very pleased to see the serious attempt to understand that you arebringing to my book. I thank you for the compliment.
---------------------------------
Bill Leach (940720.1925) --
Good observations on process control.
> ... engineered control systems don't really have goals in the samesense that living control systems do (indeed, their goals ARE theimplementation of the designers goals).
Language difficulties again. In common parlance, a goal is somethingexternal to a system; If I say I have the goal of going to the movies, themovie theater is taken to be my goal. Another usage is to treat a goal as thefinal state in which you want something else to be, out there in theenvironment. You use the latter sense above: the goal of the control system isthe designer's goal for what the control system should do.
What these usages overlook is that for any goal, the real goal is to_perceive_ that something has happened or is in a particular state. Thecontrol-system designer wouldn't be very satisfied if the control systemactually did what he had in mind, but he was unable to find out what itactually did.
The concept of the reference signal gives us a general-purpose definitionof a goal that makes sense in all circumstances. The reference signal is anexample of a perceptual signal when it is in a particular state. So thereference signal is, physically, the goal.
That definition fits both the designer and the control system. If thecontrol system has a reference signal, it has a goal for one of its perceptions(sensor signals). If the designer has a reference-signal specifying what thecontrol system is to be perceived doing, then the designer has a goal, too. Inboth cases, the goal is located physically inside the controlling system, notin its environment.
Accomplishing this change of understanding, unfortunately, requires goingall the way in absorbing PCT. It's all perception. When you look at a spot inthe service court across the net, you can see where you want the ball to go:the goal is right there, in the far left corner. So it really seems that thegoal is outside you. But as a PCTer, your understanding kicks in and says "Aha,the far left corner is a perception, represented by a neural signal in mybrain. That's how the neural signals representing the far left corner look. Somy goal is to have a perception of a ball bouncing right there where thatperception is; I'm imagining a ball doing that (with or without vividvideo)."
By that time, of course, the umpire has called you for delay ofgame.
We have to live with the fact that language incorporates a lot of theoriesthat we would really not want to support any more, but becomes awkward when wetry to speak correctly. It's just a lot easier to say "Look at the far leftcorner" than it is to say "Bring the far left corner to the center of yourperceptual field."
-------------------------------
RE: your comments on disturbances (see Mary's post, appended at end, for anifty example inspired by your mention of baking a cake):
One experiment I have done with the tracking experiments is to actuallydisplay a measure of the disturbance being applied to the cursor. Normally, thedisturbance is simply added, numerically, to the number representing handleposition and the sum determines cursor position. There's no way to separate outthe effect of the disturbance. But in this experiment, I added another "cursor"to the screen which moved up and down in proportion to the amount ofdisturbance being applied to the cursor position. If knowledge about the_cause_ of a disturbance was helpful in controlling, then people should controlbetter when this accurate information about the disturbance wasdisplayed.
Of course they didn't. They controlled worse, unless they deliberatelyignored this added information, in which case they controlled as well as usual.The problem is that the visual angle of precise vision is quite narrow, only acouple of degrees, and having to look at the disturbance indication required,if not moving the eyes, at least attending a little off center. Or at least itrequired some additional perceptual processing. Whatever the reason, thedisplay of the disturbance amplitude turned out to be a distraction, not ahelp, and control became worse.
This experiment could probably be refined so the information about thedisturbance could be obtained from a place located ON the target or the cursor,somehow. But I would still predict no better performance than when theinformation is missing. A control system doesn't use information about thedisturbance if it has information about the controlled variable. I suspect thatthe cases where sensing the disturbance proves useful are those in whichcontrol is pretty poor to begin with.
Best, Bill P.
Mary's post follows:
-----------------------------------
[from Mary Powers 940721}
Tom's tracking experiments do distinguish disturbances from variablesaffected by the disturbance. Here's another example, besides cats and dogs,from real life: cake baking.
At 6890 feet (where we live) atmospheric pressure is a lot less than at sealevel. This is the disturbance, which we are not equipped to sense (I supposewe could, but what kitchen is equipped to do so?) OR control (unless we installthe kitchen in a hyperbaric chamber :-)). Low pressure does strange things tocakes, which we can control, rather elaborately, by a) decreasing the amount ofbaking powder and sugar in the recipe, b) increasing the amount of flour andliquid (which includes eggs) and c) upping the oven temperature by 25 degrees.None of these changes controls the disturbance of low atmospheric pressure, butrather the effect of it on the physics and chemistry of the ingredients, whichin turn affect the variables actually being controlled: doneness, texture, andso forth.
Mary P.
Date: Thu Jul 21, 1994 2:48 pm PST
Subject: Re: Disturbance Control, Political Steering
From Tom Bourbon [940721.0853]
>Paul George 940720 17:30
>>[Rick Marken (940720.0845)] >>>Bill Leach(940719.19:12)
Bill L.
>>> Among other things, in engineered control systems, there areoften environmental conditions whose status is quite important to the controlprocess but is not controlled by the engineered system hardware.
Rick
>> I'd say that that's true of ALL control systems; the environmentalconditions of which you speak are called disturbances; they are the reason forbuilding a control system in the first place -- to protect variables from theunpredictable effects of these conditions.
Paul
> They are called disturbances _after_ the system is constructed.During analysis or design there are no controlled variables _yet_. That is thepurpose of the design - to determine what input functions, output functions,and hierarchical levels of control are needed to meet the purpose of keepingthe process stable and predictable.
Rick:
>> This estimate [TB: of the disturbances an engineered system mightencounter] is not really part of the control process; it is just anothervariable in the environment
Paul:
> Here is the major terminology conflict. To us 'control system' ismore than the controller. It includes the equipment, sensors, and multiplecontrollers. (so is a 'satellite control system') It even sometimes includesthe operator via interaction with displays as well as instruments or controlsnot hooked to controller input or output functions. In biology one candistinguish between the nervous system and everything else.
Yes.
> You could also draw a system boundary at the skin.
Yes, but that might not always be the best place to draw the line -- see mypost (From Tom Bourbon [940720.1653]) in which I talked about cats, dogs andpursuit tracking. The skin is just something that gets dragged along when theskeleton is moved by skeletal muscles, all as part of the incidentalenvironmental consequences of a nervous system controlling its own perceptualsignals.
> A process control system is more like a work party than aperson.
Agreed, Paul. That is the point I was trying to make when you appeared oncsg-l. I didn't do a very good job, though. When you first appeared, yousuggested that the work of Albus might be informative to PCT modelers, in thatAlbus described systems far more sophisticated than PCT models as youunderstood them then. What I tried to say in reply was that the systems Albusdescribed were not individual living control systems, but were widelydistributed groups of systems of some other kind(s). I'm happy to see that wehave all worked our ways past some of the early "heat" in the conversations tosome mutual acknowledgements that Albus-like process control is quite differentfrom perceptual control by individual living control systems. :-))
Later, Tom
Date: Thu Jul 21, 1994 2:53 pm PST
Subject: Re: B:CP impressions - 1st cut
[Paul George 940721 11:00] >[Bill Leach 940720.23:22 EST(EDT)]
Me:
>> It can be summarized thusly: All action and sensation is producedvia the interaction of neurons ... ... Thus, by default we presumeit is the mechanism used at all levels.
> I think "them's fighten' words" :-) PCT is about the idea thatBehavior is the control of perception.
I _knew_ I was going to get zapped for that :-}.
To me there are 2 significant factors in PCT, while y'all seem to focus onone. The first is that you can demonstrate that behavior is a hierarchicalcontrol process and that a neural mechanism can produce it. The second is thatperception is being controlled rather than the environment or actions upon theenvironment. I took the latter en passant, as should anyone with anyfamiliarity with epistemology.
However note that even control engineers do not view that which is beingcorrected as that which is being controlled. From _their point of view_ thecontroller is controlling the process, i.e. trying to control the environment.Of course the only way we know what is going on in the environment or with ouractions (which you normally consider part of it), is through our senses whichare abstracted up into perceptions through a hierarchy (or series, or network)of input functions. Engineers would grant that the control loop and hierarchyfunctions as described in PCT.
I grant that this control distinction is a major sticking point between youand other psychologists, but I am not sure that the first point is not the mostsignificant in terms of being able to analyze human and social behavior. Thesecond point seems to be more one of terminology or point of view than of realoverriding significance. But then I'm not a psychologist or behavioralresearcher :-).
>> When I talk about a system or control system I am usually lookingat more than one entity.
> PCT does not have a problem in dealing with multiple controlsystems.
I meant the same thing as you did in your post on process control. We tendto draw the boundary of 'a system' in a different place that PCT'rs usually do.The theory of course has no problem with distribution of control nodes orsystems, other than the mechanism of communication would possibly vary from thephysiological model.
> I should like an example of this. This sounds a bit like theargument that it is impossible to write a program in assembler that is asefficient and effective as one written using an optimizing high levelcompiler.
I intended it as more the other way around. It is more efficient to usehigh level graphic design systems than to hack code, which is much moreefficient than bit twiddling. (this is after all my field of expertise). I amreminded of a 'programmer' on one project who always hard coded incrementingloop counters, because he had never heard of a fortran 'do loop' (and neverthought to look in a manual). I know there are examples in electronics of moresophisticated circuits replacing sets of simpler ones, but can't think of aspecific on the spur of the moment.
> What do you mean by this? Higher level engineered control structuresdon't have propagation and processing lags?
No, but intra-node communication can be faster than inter-node, due tophysical distance if nothing else. An IC based CPU is a might faster (and morereliable) than the equivalent made of tubes or transistors. As a team grows insize the increased communication and coordination overhead makes the increasein capacity diminish on the order of np^2 (a.k.a an NP complete problem).Centralizing computation or logic minimizes communication at the expense ofcomplexity. All I am saying is that in a given situation a 'logic chain' andsingle control loop may be more efficient than a hierarchy of controlloops.
Date: Thu Jul 21, 1994 11:28 pm PST
Subject: Re: B:CP impressions - 1st cut
<[Bill Leach 940721.20:16 EST(EDT)] >[Paul George 94072111:00]
> I _knew_ I was going to get zapped for that :-}.
Then why d'ja do it? Huh, Huh? :-)
> To me there are 2 significant factors in PCT, while y'all seem tofocus on one. The first is that you can demonstrate that behavior is ahierarchical control process and that a neural mechanism can produce it. Thesecond is that perception is being controlled rather than the environment oractions upon the environment. I took the latter en passant, as should anyonewith any familiarity with epistemology.
There are really three that you have listed. The necessity fordemonstrating that the neural structure for control exists is really "someoneelses bag." OTOH, if what you mean by "and that a neural mechanism can produceit." is that "we" can show that biological beings exhibit control systembehavior then that is another matter altogether.
The real question there is "Can anyone show any other method of operationfor living entities?" PCT can an does demonstrate that it can model to a highdegree of accuracy the operation of living systems in limited spheres. What isimportant here with the phrase "limited" is that the evidence points to thelimited ability to model rather than limits upon a particular model match(human in this case) behavior.
At risk of being slapped into line by Bill P., I am going to propose someoverall "driving concepts" for PCT:
1. The macro view of the activities of mankind over as long a history asdetailed information is available clearly demonstrate "closed loop negativefeedback control system operation".
This single observation, I gather, was one of the overwhelming influencesupon Bill Powers concerning the possible "nature of man." You can probably addthe scientist/engineers' disdain for the "mushy" opinion based existingbehavioral "sciences".
2. Actual attempts at modeling aspects of human behavior were found tocorrelate well with real human behavior in the conduct of simple experiments.(and if I remember the story correctly, an error in prediction made by thosediscussing PCT was discovered but BOTH the human and the model behaved in thesame fashion -- thus interpretation of the significance of control theory waswhat was in error, not the theory itself).
3. There seems to be physical evidence of the existence of actual controlsystems in living systems (including, of course, people). The discovery ofactual control system loops is "encouraging" but is not central to the theory.If the biologists had not found such structures, the rest of the behavioral"sciences" would still have the problem of explaining closed loop controlsystem behavior some other way and none has been demonstrated to work.
> I grant that this control distinction is a major sticking pointbetween you and other psychologists, but I am not sure that the first point isnot the most significant in terms of being able to analyze human and socialbehavior. The second point seems to be more one of terminology or point ofview than of real overriding significance. But then I'm not a psychologist orbehavioral researcher :-).
If you reword that first line to "... is THE major ...", I believe that youwill have phrased it correctly.
As I tried to indicate above, the physical evidence of structure is not atissue and would not be even if it appeared to be conflicting. If the biologicalevidence indicated something other than closed loop control systems, theirwould still be the problem of explaining how something that is not a closedloop control system behaves as though it is one.
>> PCT does not have a problem in dealing with multiple controlsystems.
> I meant the same thing as you did in your post on process control. Wetend to draw the boundary of 'a system' in a different place that PCT'rsusually do. The theory of course has no problem with distribution of controlnodes or systems, other than the mechanism of communication would possiblyvary from the physiological model.
Indeed, if I have learned anything from Tom Bourbon at all, it is that"social aspects" of human behavior are explainable by analyzing the dynamics ofcontrol systems operating in a common environment.
>> I should like an example of this. This sounds a bit like theargument that it is impossible to write a program in assembler that is asefficient and effective as one written using an optimizing high levelcompiler.
> I intended it as more the other way around. It is more efficient touse high level graphic design systems than to hack code, which is much moreefficient than bit twiddling. (this is after all my field of expertise). I amreminded of a 'programmer' on one project who always hard coded incrementingloop counters, because he had never heard of a fortran 'do loop' (and neverthought to look in a manual). I know there are examples in electronics of moresophisticated circuits replacing sets of simpler ones, but can't think of aspecific on the spur of the moment.
It is (generally) a more efficient use of resources including time, toprogram in a high level language. However, in anything but a trivial program itis always possible that assembly code could be written that would be moreefficient as a program than that generated by a compiler. That programs havereached levels of complexity that would make it not only impractical to makethe attempt but would be physically impossible for even a medium sizedprogramming team to do so.
The statement is true because of the very nature of programming and logicprocessing. Indeed, high level languages are nothing more than programs writtento convert tokens and symbols to binary code but such conversions are somewhatgeneral. You might want to take a look at some of the comparisons that havebeen made between code size and function as a function of the method ofgeneration.
In terms of efficiency, computers have moved away from efficiency at aremarkable rate... it is just that processing power has more than kept pace(most of the time).
I am "overly belaboring" the point here except that I think that it isworthwhile to consider that what we often perceive as an improvement inengineered control systems is in fact a system with vastly greater processingpower that is only able to make marginal improvements in control systemcapability. The "real" improvement is in the change to the environment that theDESIGNER is working in. People do things in control system design today as amatter of course that would not have been much more than dreamed of even asrecently as 10 years ago. The reason that they would not have done these thingswas not because they were impossible but rather because they were impractical(I will admit, of course, that there are things that are done today that were,at least perceived, to be impossible at one time).
>> What do you mean by this? Higher level engineered controlstructures don't have propagation and processing lags?
> No, but intra-node communication can be faster than inter-node, dueto physical distance if nothing else.
For the technology in actual use. Indeed, we have deployed communicationssystems whose throughput would "crush" the sort of controllers that you aretalking about but in general it is not practical to use such communicationssystems for your application.
> An IC based CPU is a might faster (and more reliable) than theequivalent made of tubes or transistors.
I'm not sure of the significance of this statement... is your company stillusing tubes? :-)
> As a team grows in size the increased communication and coordinationoverhead makes the increase in capacity diminish on the order of np^2 (a.k.aan NP complete problem). Centralizing computation or logic minimizescommunication at the expense of complexity. All I am saying is that in a givensituation a 'logic chain' and single control loop may be more efficient than ahierarchy of control loops.
This is true but its' relevance is dependent upon how control isstructured. A biological control hierarchy seems to be an example in theextreme for what you are saying is NOT a good control methodology. If you thinkyou have ever seen a "large" distributed control system design, just ponder fora moment the sheer number of "controllers" that a human might have.
The "trick" there is that each level also "removes" raw detail in aperceptual signal so that each "neural signal" in successively higher levelsrepresents a greater amount of information. Likewise, each high level reference"decomposes" into many lower level references on the way down. It is sort oflike the idea that a node in a binary tree represents all of the nodes below it(in its' chain). The datum is (at least close) to the same size and duration asdatum "at the skin". Thus, no NP problem).
-bill
Date: Fri, 22 Jul 1994 07:14:12 -0600
Subject: control of perception
[From Bill Powers (940722.0510 MDT)] Paul George (940721.1100)--
> However note that even control engineers do not view that which isbeing corrected as that which is being controlled. From _their point of view_the controller is controlling the process, i.e. trying to control theenvironment.
I am beginning to think that this is how _naive_ engineers view theprocess. Of course their task is to be sure that the environment is controlled,because that is what the customer wants. But the control system itself, the onethey build and that must operate on its own, can control only what it perceivesof the environment. If the perceptual system is set up to perceive the wrongthing, the engineer may believe that the environment is being properlycontrolled, when in fact the control system is controlling some other aspectsof it which only happens, for now, to keep the variables the engineer isinterested in within the required limits. But a change in circumstances canleave the control system controlling quite successfully while the environmentgoes to pot.
Suppose the engineer wants to control the temperature of a bath. He givesthe control system a thermocouple to register the temperature, constructs therest of the loop, and all is well. But the thermocouple, as it happens, issensitive to radiated energy as well as conducted energy. During the tests, ithappened to be shielded from radiated energy, or there were no radiant energysources to worry about. So everything worked. But when the control system ismoved into the factory where there are powerful lights overhead, or sunlightcoming in through windows, it will start controlling for the sum of the inputenergies from ALL sources, not just the intended one. When the lights go on,the temperature in the controlled bath will go down!
Worse yet, when the engineer is summoned to see this problem, he goes rightup to the bath and stands there, and sees that everything is operatingnormally. When he goes away, removing his shadow from the thermocouple, downgoes the bath temperature again.
Engineers may intend to control variables in the environment, but thecontrol systems they build can control only their own sensory signals -- nomatter how the engineers think about it. Smart engineers know this,
if only intuitively; this is why they specify periodic calibrations of thesensors. When a sensor drifts in calibration, the control system continues tohold the sensor's output signal at the set point, but this results in theexternal variable's changing. This is why they make sure that a sensor whichnominally senses one variable, like light intensity, is not also affected byother variables, such as magnetic fields or temperature or humidity.
A control system controls its own sensory signal. This is not a matter ofinterpretation; it's how control works.
Best to all, Bill P.
Date: Fri, 22 Jul 1994 11:11:15 EDT
Subject: Re: control of perception
[Paul George 940722 11:20] >[Bill Powers (940722.0510 MDT)]
> I am beginning to think that this is how _naive_ engineers view theprocess.
Note I meant that there is no disagreement on how a control loop actuallyworks, just on focus on what is spoken of being 'under control'. However, Ithink something more like 'blissfully ignorant' would describe a frighteninglylarge number of them. Unfortunately a sense of infallibility and of order inthe universe is a common attribute of engineers. Hazard analysis is rarelyperformed, and there are a lot of naive assumptions about the environment andequipment.
> If the perceptual system is set up to perceive the wrong thing, theengineer may believe that the environment is being properly controlled, ...But a change in circumstances can leave the control system controlling quitesuccessfully while the environment goes to pot.
That is indeed the challenge in process control system design; figuring outwhat needs to be sensed and how to sense it.
> Suppose the engineer wants to control the temperature of a bath. Hegives the control system a thermocouple to register the temperature,constructs the rest of the loop, and all is well. But the thermocouple, as ithappens, is sensitive to radiated energy as well as conducted energy. Duringthe tests, it happened to be shielded from radiated energy, or there were noradiant energy sources to worry about. So everything worked. But when thecontrol system is moved into the factory where there are powerful lightsoverhead, or sunlight coming in through windows, it will start controlling forthe sum of the input energies from ALL sources, not just the intended one.When the lights go on, the temperature in the controlled bath will go down!Worse yet, when the engineer is summoned to see this problem, he goes right upto the bath and stands there, and sees that everything is operating normally.When he goes away, removing his shadow from the thermocouple, down goes thebath temperature again.
I wonder if you have any idea how realistic this example is? (though RF,magnetic, or electrical interference is more common) Unfortunately a lot ofsystems are designed this way. Problems like this often show up in the field.And funny thing, when the equipment goes back to the shop, there is no problemfound.
"...And he examined the test data and saw that some of it was good, andsome not so good. He therefor divided the good data from the bad and he calledone 'results' and the other he called 'spurious anomalies'+ ;-)
Paul
Date: Fri, 22 Jul 1994 10:28:06 EDT
Subject: Re: B:CP impressions - 1st cut
[Paul George 940722 10:20] >[Bill Leach 940721.20:16)]
> OTOH, if what you mean by "and that a neural mechanism can produceit." is that "we" can show that biological beings exhibit control systembehavior then that is another matter altogether.
Kind of both, if the B:CP discussion of spinal reflexes is correct (whichto my level of neurological knowledge it is). The latter point is what I meantby:
>> The first is that you can demonstrate that behavior is ahierarchical control process...
> If you reword that first line to "... is THE major ...", I believethat you will have phrased it correctly.
:-) Actually, I did phrase it that way the first time and later toned itdown out of politeness.
> As I tried to indicate above, the physical evidence of structure isnot at issue and would not be even if it appeared to be conflicting. If theillogical evidence indicated something other than closed loop control systems,their would still be the problem of explaining how something that is not aclosed loop control system behaves as though it is one.
Want to run that one by me again? You are saying that if I could prove thatthe CNS anatomically did not use closed loop control structures that the burdenof proof would be on me to show how and why it worked?? From my point of viewit would simply show PCT to be incorrect in asserting that closed loop controlis the mechanism for organismic behavior. It would still be useful foranalyzing or predicting it. Two different mathematical functions that give thesame results within a given range are equivalent within that range.
> A biological control hierarchy seems to be an example in the extremefor what you are saying is NOT a good control methodology........ It is sortof like the idea that a node in a binary tree represents all of the nodesbelow it (in its' chain). The datum is (at least close) to the same size andduration as datum "at the skin". Thus, no NP problem).
I'm not saying it is bad, just not always the most efficient or simple (interms of number of components). The situation I am talking about is where thereare a large number of nodes on the same level which must interact in order to'evaluate' (not necessarily control) enough perceptions to 'make decisions','recognize patterns', or make plans. Computers use memory and logic structuresother than trees (particularly in AI) for a reason. The levels 1-3 of abiological system are clearly hierarchical anatomically . Hierarchicalstructures to filter, summarize, or compress information are common. However,the hierarchy is often in terms of 'strata' rather than levels. As Bill Bindicated in his response to my post:
>[Bill Powers (940721.0815 MDT)]
> I keep thinking of cases where what I call a third-level controlsystems seems to operate via a fifth-level control system, and so on. So farI've been able to resolve most of these possibilities, but the question isstill open whether we should think of these "levels" as "dimensions" instead.... I opted for the hierarchy, but have wondered ever since whether thereisn't something to the other view, too.
At some point the 'canonical model' will likely break down due to theexplosion of the number of nodes required. It takes a lot of control loops tocontrol something like writing this post. When you get to the level of thesensory processors and upper cortex functions such as language,I stronglysuspect that network rather than hierarchical structures predominate in orderto generate the constructs we usually term 'perceptions' or 'ideas'. There willbe more 'horizontal' connections than 'vertical'. I also suspect that controlfunctions will involve more 'algorithmic' processing. However, I am not surehow you could tell empirically, though the stuff being done with PET scannersmight help.
Paul
Date: Sat, 23 Jul 1994 17:07:49 -0400
Subject: Re: B:CP impressions - 1st cut
[Bill Leach 940723.16:23 EST(EDT)] >[Paul George 940722 10:20]
>> As I tried to indicate above, the physical evidence of structureis not
> Want to run that one by me again? You are saying that if I couldprove that the CNS anatomically did not use closed loop control structuresthat the burden of proof would be on me to show how and why it worked??
First, let me say that this was my opinion and not necessarily shared byanyone else here. In particular, I know that Bill P. HAS studied anatomy to tryto look for evidence in either direction and I'm sure that he cares very muchwhat sort of new evidence shows up.
In a sense yes. The issue with PCT as I understand it IS NOT therepresentation of the structure, the possible levels and hierarchy. The issueis closed loop negative feedback control system is the only known mode ofoperation of living systems that explains what is actually seen.
That PCT explains S-R behavior is not even the point though the fact thatit explains exceptions to S-R behavior is getting closer.
If you "come up with proof" that a person's neurological structure CAN NOTbe a control system, then yes you indeed have a great burden of proof on yourhands -- how does a non-control system consistently exhibit control systembehavior?
> From my point of view it would simply show PCT to be incorrect inasserting that closed loop control is the mechanism for organismic behavior.It would still be useful for analyzing or predicting it. Two differentmathematical functions that give the same results within a given range areequivalent within that range.
I don't challenge the last sentence of your statement there at all. Show methe function that provides the equivalent results. Short of metaphysics thereis no known phenomenon that can make a non-control system behave exactly like acontrol system. If you do find such a phenomenon maybe the second sentencewould still be true.
> I'm not saying it is bad, just not always the most efficient orsimple (in terms of number of components).
I agree and will even add that when enough is known about the limits ofenvironmental disturbances possible, even the choice of controlled perceptionsmay be poor.
> At some point the 'canonical model' will likely break down due to theexplosion of the number of nodes required.
No, I don't think so. Even the discussions and models currently existing inthe PCT world do major consolidation of control loops and even hierarchy. Wheresuch "would break down" is in any attempt to exactly model the structure of theliving being as opposed to modeling its behavior for specific instances ofcontrol.
-bill
Date: Sat, 23 Jul 1994 16:53:04 -0600
Subject: Levels and Misc
[From Bill Powers (940723.1540)] Bill Leach (940723.14:41)
(your note to Paul George):
> The issue with PCT as I understand it IS NOT the representation ofthe structure, the possible levels and hierarchy. The issue is closed loopnegative feedback control system is the only known mode of operation of livingsystems that explains what is actually seen.
There's no way to separate these issues. A conceptually neat model thatviolates known neuroanatomy would be just as useless as a neuranatomicallycorrect model that blew up when you tried to run it. What we must have is aworkable model that recreates the behavior we actually observe, and fits whatwe know of the nervous system and muscles as well -- and works in the worlddescribed by physics and chemistry. This is supposed to be a model of a realsystem, not an intellectual exercise.
--------------------------------------
RE: boundaries.
The basic rule is: put yourself inside the system you're talking about. Ifyou're the dog, your perceptual field contains a cat moving unpredictablyaround, as well as your nose and paws -- but no dog. If you're the cat, there'sa dog in your perceptual field along with your nose and paws, but no cat. As anexternal observer, you are looking out of the wrong boundary. You can see a catAND a dog, and the noses you see (other than your own) are seen in profile orin front, not from behind them. That's the incorrect view for understandingeither system.
Best, Bill P.
Date: Sun, 24 Jul 1994 00:39:55 -0400
Subject: Re: Levels and Misc
[Bill Leach 940724.00:17)] >[Bill Powers (940723.1540)]
> (your note to Paul George):
>> The issue with PCT as I understand it IS NOT the representation ofthe structure, ...
> There's no way to separate these issues. A conceptually neat modelthat violates known neuroanatomy would be just as useless as a ...
I agree except that I also take the position that a "known" neuroanatomythat denies control loop operation would itself be wrong. I guess, in essence Iam taking the stand that "so called" proof that the system is not a controlsystem would be ok if and only if, it could also show that the control systembehavior that is actually observed could also be explained.
> What we must have is a workable model that recreates the behavior weactually observe, and fits what we know of the nervous system and muscles aswell -- and works in the world described by physics and chemistry. This issupposed to be a model of a real system, not an intellectual exercise.
I agree with this but what I am saying is that any other assertion aboutthe structure MUST also account for the behavior. What "we know" about thenervous system is pretty tentative at best and a "generally accepted" theory ofthe nervous system structure that was used to assert that control systembehavior had to be an illusion would not be particularly convincing to me atthis point. Science has refuted a correct theory more than once in thepast.
> RE: boundaries.
Ok, I think that Tom straightened me out on that one. I really will try tobe more "reference oriented" to the control system itself in the future unlessspecifically stating otherwise.
-bill
Date: Thu, 4 Aug 1994 16:14:14 EDT
Subject: B:CP Reactions - Cut 2
[Paul George 940804 1600]
{Directed mostly to Bill Powers, but comments welcome from anywhere}
Having finished B:CP, I think it is a good piece of work. Too bad so muchof it seems to be ignored in most of the research and the discussions here,which seem focused on the 'worm's eye view' of PCT (1st & 2nd ordercontrol). Perhaps because the other concepts are deemed 'uninteresting'?:-)
I do have a few questions & observations:
In terms of the reorganization system, why must the reference variables be_intrinsic_ and physiological? (I equate 'intrinsic' with hardwired or inborn)Some clearly would be, for example detecting life threatening situations orphysiological conditions (Thirst, hunger, various levels of stress or pain,danger, 'instincts'). However, I don't see anything which would preclude havinglearned or otherwise generated reference levels as well. Recognizing the needto adapt would seem to be a characteristic of higher levels of intelligence. Ageneral 'error level' or 'discontent' intrinsic would help to 'awaken' thereorganization function, but seems to me unsatisfying for directing it. Noninnate values would also help explain how some people seem to change theirbehavior in self destructive ways. Their hierarchy of reorg reference variableshave placed some artificial(??) goal at a higher priority than the natural orintrinsic ones. In other cases they may be trying to reconcile two incompatiblegoals. (I'm not sure that goal is the appropriate term, but is the one I'm mostfamiliar with).
Note: I have not explicitly seen the concept of priorities betweenreference variables, perceptions, or control loops, except through thehierarchy and the actual constructs in the 'little man'. Levels of thehierarchy would appear to be more than one loop deep. I am not at all sure thatsuch a concept is truly required, but at some level some disturbances are moreimportant than others.
Re the close of chapter 14 on Platt's work. What experimental work has beendone in the last 20 years on reorganization? Personally the application ofcontrol theory to adaptation, learning, and communication was what sucked meinto this group, combined with it's effects on interpersonal relationships. Thebook has a number of references to work that needed to be done, but it isunclear to me as to what actually has happened.
I personally think you do your 'cause' a bit of harm by splitting off thereorganization hierarchy from the central control hierarchy, and then ignoringit in most of your posts and articles. I would publish figure 14.1 a littlemore widely. Similarly, just pushing the simple canonical control diagram asthe definitive model instead of the figure 15.3 'final form' is misleading andI think hurts acceptance of your ideas. You don't highlight that the canonicalmodel is a simplification of the 'true' form which you use because it issufficient for your experimental purposes.
A large part of my initial impressions of weakness or oversimplicity inHPCT was because you (particularly Rick & Tom) appeared to be saying moreelaborate structures were unneeded. I recognize that most PCT research isfocused on levels 1-3 of the perceptual control hierarchy, partially due toamenability to modeling, but creates the impression that it is all HPCTconsists of. The simple perceptual control loop does not explain all of what iscommonly called behavior, much less psychology. The reorganization controlhierarchy, memory,and the perceptual & memory switches are necessaryconstructs to deal with common concepts of planning, imagination, etc. Furtherthe levels of the PC hierarchy above 3 are rarely explained, other than byoccasional asides - e.g. "The lower six levels are concerned with control ofintensities, sensations, configurations, transitions, events, andrelationships. "
[Bill Powers (940803.1510 MDT)]
It is not intuitively obvious that these constructs derive from the simpleconcept of control. The terms are also subject to variableinterpretation.
As the [From Bill Powers (940803.1510 MDT)] exchange with comp.aiindicates, your critics and the uninitiated dismiss you because you do notpublicize the 'non-core' ideas, except through casual asides (in myexperience), and so they presume such concepts are not incorporated inPCT.
Please accept a suggestion from a newbee who recently went through tryingto get up to speed with PCT, even given a fairly strong understanding ofcontrol. Given the paucity of 'public domain' sources, I would recommendwriting a little white paper and incorporating it in the monthly post (as wellas archiving it (you could really use a FAQ). Few people are willing to shellout a significant amount of money or expend a lot of effort to investigate anapparently fringe or 'crackpot' (ITHO) theory.
I would summarize the levels of control discussed in chapters 7-13 from arole/responsibility standpoint at about a paragraph apiece. The introductorypages and chapter summaries could provide most of the meat. Introduce the reorghierarchy as well, possibly with an elaboration of figure 14.1 (I think thishides the concept of their being two 'planes' of control hierarchies:x/z=PC,y/z=RC). This heads off objections concerning higher level behavior(from the standpoint of internal perceptions) and learning.
Next provide a paragraph for each of the neural gate constructs for theunderpinnings of the feedback control mechanism and logic functions usedelsewhere. Use the spinal reflex model (say figs 7.3 or 7.4) to discuss thebasic control phenomena.(this is the part that is well covered by the currentintro material, but it's derivation from physiological constructs is nothighlighted). I would deem this a key point for acceptance. Use 'the parable ofthe rubber band' experiment (16.3), particularly the coin extension(Brilliant!!), to further illustrate the power of the concept.
Next briefly explain the final form diagram (15.3) with the memory and the4 'control modes' (I think this was my favorite part of the book. What a neatmodel!).
The organization could perhaps use work, but I think it can be done clearlyin less than 10 pages{Shorter than many of your posts ;-)} Also, considerputting your response to Bruce Buchanan (or perhaps both posts) into the intromaterials. I wish I had read it much earlier as it would have avoided muchmisinterpretation as to what PCT encompassed. I am happy to see that PCT doesin fact cover the areas I had anticipated it should.
Continuing to enjoy the discussion, Paul
Date: Fri, 5 Aug 1994 13:50:09 EST
Subject: worm's eye view
[Avery Andrews 940805.1347] (Paul George 940804 1600)
> Having finished B:CP, I think it is a good piece of work. Too bad somuch of it seems to be ignored in most of the research and the discussionshere, which seem focused on the 'worm's eye view' of PCT (1st & 2nd ordercontrol). Perhaps because the other concepts are deemed 'uninteresting'?:-)
I won't speak for the others, but I've been focussing on the lower levels(so far) because we have a long tradition of investigation of the upper levels(going back to Aristotle and the Stoics, if not further), and large numbers ofclever and well-funded people still beavering away at them, but it all is andwill remain very up in the air until we understand how the upper levels cashout as actual activity, which requires connecting them to lower-level systems,which requires reaching some minimal level of understanding of how thesework.
Plus it seems to me the that actual properties of even simple closed-loopsystems are often seriously misunderstood by people (such as Fowler and Turvey,or Abbs and Winstein) who one would really expect to understand them properly,so there is obviously some kind of problem there.
There is also the fact that the behavior of control systems is often wildlycounter to even well-informed intuition, so that if you can't model, you run abig risk of producing bullshit. But our ability to model is limited by, amongother things, our ability to construct perceptual functions, which is,especially at the higher levels, pretty minimal.
Avery.Andrews@anu.edu.au
Date: Fri, 5 Aug 1994 09:51:52 -0700
Subject: Appealing to Complexity Worshippers
[From Rick Marken (940508.0945)] Paul George (940804 1600)
> Having finished B:CP, I think it is a good piece of work.
To me, this is a bit like calling Newton's "Principia" a "good piece ofwork". But I guess I have to agree with your basic assessment.
> Too bad so much of it seems to be ignored in most of the research andthe discussions here, which seem focused on the 'worm's eye view' of PCT (1st& 2nd order control). Perhaps because the other concepts are deemed'uninteresting'? :-)
I think Avery Andrews (940805.1347) gave an EXCELLENT reply to this. Inparticular, I like his observation that the "actual properties of even simpleclosed-loop systems are often seriously misunderstood by people". I find itamusing when people storm off to model the higher levels before they grasp eventhe most basic concepts of PCT, such as the nature of control, the control ofperception, testing for controlled variables, etc. By ignoring the basics,these people fail to see that a great deal of what might SEEM like "higherlevel" behavior is just the side-effects of very simple controlling (forexample, consider the behavior of the simple control systems in the CROWDprogram or of the interacting humans in Tom's experiments on cooperation andconflict).
> A large part of my initial impressions of weakness or oversimplicityin HPCT was because you (particularly Rick & Tom) appeared to be sayingmore elaborate structures were unneeded.
In PCT, we try to explain phenomena with as simple a model as possible. Itseems to me that one of the diseases of modern behavioral science is thefascination with complexity per se. I'll call it " complexity worship". Itseems that many behavioral scientists think a model is weak and/oroversimplified if it is not complex (this is a big change from the scientificgoals of Newton and Einstein). These behavioral scientists seem to becontrolling for a high level of perceived complexity as a means of perceivingthemselves as understanding behavior.
The modern disease of "complexity worship" seems to stem from threeproblems that are endemic to the behavioral sciences:
1) The first is an addition to the superficial; behavioral scientists arefascinated (as we all are) by the visible side effects of controlling, whichcan look quite complex.
2) The second is an addition to the verbal labels used to refer to theseside effects. Thus, if a collection of observable side effects is called"optimal trajectory selection" it seems like one is dealing with something agreat deal more complex and "high level" than "pressing a bar". Many PCTstudies that had seemed simple and "low level" suddenly become "important andrelevant" when they are given names like "helping", "cooperation", "conflict","leadership", "learning", etc.
3) The third is an apparent aversion to experimental test and theacceptance of incredibly poor data when such tests are performed. There seemsto be a growing interest in theory qua theory; there is very little testing ofsuch theories against data. And the theories that are developed are not reallyattempts to account for real data; they are attempts to account for verbal_descriptions_ of data. So we see people trying to come up with theories of"path planning", "alerting", "intelligent search", etc. That is, people aretrying to draw diagrams that they imagine will produce the kind of behaviorthat they imagine people would describe with these words.
People who enjoy PCT tend to be those who find that their understanding ofhuman nature is controlled by building working models (the simpler, the better)that produce behavior that exactly duplicates real data. People who enjoy PCTtend to be "phenomena freaks"; they are interested in understanding a realphenomenon that is really interesting that they can really experience inthemselves and others - - CONTROL. People who enjoy PCT tend to enjoy simpleexplanations (probably because we a simple minded) that are powerful (theyexplain a LOT of data); basic PCT explains a LOT of data.
I don't think there is much that PCT can do for behavioral scientists whosuffer from the three addictions I listed about -- to the superficial, to theverbal, to the theoretical (undisciplined by the phenomenal -- i.e. data). Butdon't blame PCT for the fact that most of the behavioral science community isclueless. PCT only seems to be "simple minded" if you approach it from theperspective of clueless behavioral science.
In fact, the basic PCT model has a great deal to offer people interested inhigh level, "real life" problems; the discussions at the recent CSG meetingwere evidence of that. Once you understand the basic PCT model (including thenotion of hierarchy and reorganization) you can cut through a lot of theapparent complexity of behavior to see what people are actually doing(controlling perceptions) and why they often have trouble doing what they want(conflict).
In order to "sell" PCT to the "complexity worshippers" in the behavioralsciences, we would have to make it seem like PCT gives them what they want --descriptions of the superficial complexities of behavior, understanding throughverbalization and theoretical complexity for its own sake. I, personally, wouldrather just stay in the "simple minded" PCT ghetto, with simple-minded friends(like Tom, Bill and Avery) doing simple minded research and publishing it insimple minded books like "Simple MINDed READINGS".
Simply, Rick
Date: Fri, 5 Aug 1994 13:29:27 -0600
Subject: Re: B:CP evaluation, part 2
[From Bill Powers (940805.1150)] Paul George (940804.1600)
> Having finished B:CP, I think it is a good piece of work. Too bad somuch of it seems to be ignored in most of the research and the discussionshere, which seem focused on the 'worm's eye view' of PCT (1st & 2nd ordercontrol). Perhaps because the other concepts are deemed 'uninteresting'?:-)
Having finished the book, you can now speak with more assurance about whatPCT is than others who haven't -- as you now realize.
Avery Andrews explained the position well. There are lots of peopleguessing about how the higher levels work, but there's no unifying principle totie such work into the whole system. The biggest problem is in designingresearch and doing simulations without understanding how higher-levelperceptions work, or even exactly what the higher levels are. So we work wherewe're more sure of our ground, trying to develop methods that will applygenerally but not going too far beyond what we can demonstrate. It will be upto smart people like you to push the boundaries further upward in thehierarchy.
> In terms of the reorganization system, why must the referencevariables be _intrinsic_ and physiological? (I equate 'intrinsic' withhardwired or inborn).
The main reason is that I needed something that _could_ be inborn, toprovide the basis for the development of the hierarchy from infancy (or before)onward. I see the hierarchy in the neonate as a set of possible controlsystems, with the necessary sorts of computations provided for at each levelbut no wiring and no organization yet. We have to learn to perceive everything,at every level, and also how to control what we perceive. As I said in thebook, this is a worst-case model, where there is an absolute minimum of inbornorganization. The reorganizing system has to work under these worst-caseconditions, where even the _concept_ of an algorithm or systematic strategyhasn't been developed yet.
If everything from sensations to systems concepts has to be learned, weobviously can't count on anything in those categories as an aid to learning.The process of reorganization, in short, can't be intelligent, because it hasto work properly before any intelligence is developed. So I asked myself whatthe basis for reorganization could possibly be with those rather severerestrictions, and came up with the answer that it has to be concerned with thestatus of the organism itself. That's where the idea of intrinsic variablescame from (that, and from W. Ross Ashby, who called these "critical variables"and postulated a random switching device which was the direct ancestor of myidea). There's no real need to guess at what the intrinsic variables are; allwe need to specify is that they be affected by the way the organism behaves inan environment, and that no knowledge of the external world or its laws beinvolved.
Ashby assumed that there were built-in upper and lower limits on criticalvariables, exceeding which started the random switching process. I substitutedthe idea of reference signals, but that's a minor difference (two-way insteadof one-way control). He also saw that such a system would be "superstable" inthat it would keep randomly switching until the critical variables were onceagain within limits -- regardless of _why_ they came back within limits. Isimply adopted that principle; it's still the basic principle of reorganizationin PCT.
I was always somewhat unsure whether random switching could really beefficient enough to accomplish real reorganization, until I heard of Koshland'swork on chemotaxis with E. coli, and set up some simulations to test his idea.The results astonished me. This turns out to be an unbelievably effective wayof controlling things in a few dimensions at once. As long as reorganization isapplied to small parts of the system at any one time, it can very quickly leadto local optimizations. The random nature of the reorganizing process evenprovides for getting out of local minima if they aren't too deep. I have shownthat this method can be used to solve 50 equations in 50 unknowns in half anhour or so.
> However, I don't see anything which would preclude having learned orotherwise generated reference levels as well. Recognizing the need to adaptwould seem to be a characteristic of higher levels of intelligence.
But you've put your finger right on the problem, haven't you? Where dothose higher levels of intelligence come from? They have to be learned, too.And if they are to be learned, they must be the product of some learningprocess that predates them. Once each new level is organized, many controlproblems that could formerly be solved only by random reorganization can now besolved by a systematic control process, which is far faster and more accurate,and can keep errors so small that intrinsic errors are never seen -- at leastfor reasons having to do with that level of control. It's even possible thatthe reorganizing system creates (at the logical level, probably) specificstrategies for learning, although it's learning of a different kind from therandom trial-and-error of reorganization.
> A general 'error level' or 'discontent' intrinsic would help to'awaken' the reorganization function, but seems to me unsatisfying fordirecting it.
Yes, general error level in the hierarchy (along with other types ofvariables inside the organism) is considered an intrinsic variable, with abuilt-in reference level of zero. When you say that reorganization isunsatisfying for "directing" the results, however, you're not recognizing thesurprising property that random reorganization CAN be directed toward achievinga specific end-state. Whatever the variable under control, reorganization isdriven by the difference between that variable and a corresponding referencestate. This is the "intrinsic error" signal. Specifically, the interval betweenreorganizations depends on the rate of change of intrinsic error times themagnitude of that error -- which amounts to the first derivative of the squareof the error. That is the model that seems to work the best.
What's hardest to grasp here is that the reorganizing system doesn't carewhat organization it produces as it acts on the system being reorganized. TheONLY thing it is concerned with is the intrinsic error. It will stopreorganizing only when the intrinsic error gets small enough. There many bebetter organizations, or the intrinsic error might drop to zero because of thearbitrary action of an external agency. None of that matters to thereorganizing system. It just wants its own intrinsic error to be zero. If it iszero, or small enough, reorganization stops, leaving _whatever_ organizationexists at that point to go on operating.
It is this extreme pragmatism of the reorganizing system that makes it soeffective. It does not need to know anything about the environment, or theorganism, or the brain, or anything else besides the states of its own inputvariables in relation to their respective reference signals. It will tryliterally anything, and at random, until intrinsic error disappears, and thenit will stop acting. How, during this process, it has changed the relationshipbetween the behaving system and its environment is completely irrelevant to it.It can't either know or care about that. That is why it can produce a workableorganization in any kind of environment. Its very stupidity is what makes it sopowerful.
RE: experimental work on reorganization.
Very little has been done with real people, although Dick Robertson hasdone some experiments in which reorganization appears to occur, and FransPlooij has recorded data on infant learning in chimpanzees and humans that seemto show definite periods of reorganization. We have done a number ofsimulations, one of them being a control system model that matches itself tobehavior of a real person by treating the difference between the model's andthe person's behavior as an intrinsic variable and randomly reorganizing theintegration factor of the model's output function to reach a minimum in thedifference. As mentioned, I've tested some models involving multiple controlsystems controlling a shared world of multiple variables, and reorganizinguntil the system of simultaneous equations is solved for independent control byeach system of a different aspect of the shared environment. That's not a lot,but such things take time to work out and at least it's something.
RE: presenting the theory
> I personally think you do your 'cause' a bit of harm by splitting offthe reorganization hierarchy from the central control hierarchy, and thenignoring it in most of your posts and articles.
Well, sheesh, how much can we talk about at once? We talk about the thingswe have done the most work with. Actually, if I had responded to your initialinquiries with a complete summary of everything that is in the PCT model, youwould have thought I was some kind of nut. Haven't you ever seen real nut mail?If somebody you had never heard of sent you a two-inch-thick packet of drawingsand text with the title, "A complete theory of everything that all organisms dounder all conditions," wouldn't you alert the nearest looney bin? Most peoplewho first hear about PCT have the impression that I thought it all up lastweek, so they're surprised that I haven't cited someone who said somethingsomewhat similar last month. The idea that there's been this whole longdevelopment going on for 40 years, and that they've never heard of it, neveroccurs to them (and why should it?). And when I can't explain it all in onebreath, they start fidgeting.
> Please accept a suggestion from a newbee who recently went throughtrying to get up to speed with PCT, even given a fairly strong understandingof control. Given the paucity of 'public domain' sources, I would recommendwriting a little white paper and incorporating it in the monthly post (as wellas archiving it (you could really use a FAQ).
I'd love to have an FAQ. How much do they cost, and can I get a used onecheaper? By the way, what is an FAQ?
I'm still leery of trying to cram the whole theory into the monthly intro,for the reasons cited above. Would you like to try it, being fresh from thelearning experience? You seem to have a pretty clear view of what might makethe whole thing more quickly understandable. Why not post a first try and letsome of the other newcomers chime in. It might turn out to be a reallyeffective document. "How I finally got the idea of PCT," or something. In fact,your suggested outline sounds pretty good to me.
Best, Bill P.
Date: Fri, 5 Aug 1994 15:57:20 EDT
Subject: Re: Appealing to Complexity Worshippers
[Paul George 940805 16:00]
>[Rick Marken (940508.0945)], also Avery Andrews (940805.1347)
> I find it amusing when people storm off to model the higher levelsbefore they grasp even the most basic concepts of PCT, such as the nature ofcontrol, the control of perception, testing for controlled variables, etc. Byignoring the basics, these people fail to see that a great deal of what mightSEEM like "higher level" behavior is just the side-effects of very simplecontrolling.
I share your reaction, and acknowledge that a little knowledge is adangerous thing. I intend to "wait for fullness". But I wish that _you people_had explored more in those areas, given that you presumably do understand.(Ofcourse I haven't read all you have written) From one standpoint all you havedone for 20 years is elaborate the evidence to support the models andexperiments described in B:CP (not to suggest that the effort was wasted ortrivial). Of course a major reason for 'slow' progress may have been thataffordable computers with sufficient power only became available in the last5-10 years.
I don't think I suffer (much) from the Three Addictions (particularly thethird). I have no real issue with what most PCTers work with. Simple models arebest, allowing for Einstein's caveat of "...as simple as possible, but nosimpler". I just personally find the concepts of memory as reference value, theswitching construct, and reorganization fascinating. Forgive me if I rely on_your_ work and conclusions modeling the basics :-). I'm mostly interested inconclusions, implications, and applications. Of course my job deals withdesigning processes and technology to allow people to adapt and communicate inorder to deal with changing environments and problem solving.
> In order to "sell" PCT to the "complexity worshippers" in thebehavioral sciences, we would have to make it seem like PCT gives them whatthey want -- descriptions of the superficial complexities of behavior,understanding through verbalization and theoretical complexity for its ownsake. I, personally, would rather just stay in the "simple minded" PCTghetto, with simple-minded friends (like Tom, Bill and Avery) doing simpleminded research and publishing it in simple minded books like "Simple MINDedREADINGS".
However, if propagation of truth or knowledge is your goal, rather than acloistered purity of thought and purpose, you need to communicate with theunenlightened. IMHO you need to first communicate that the theory can encompassthe complexities that many psychologists and commoners ;-) observe in humanbehavior, and then focus them towards the basics. Then they can be amazed athow much so called complex behavior simple hierarchical control can model. Butfirst you have to catch the mule's attention.
If you make it easy for them to classify you as monomaniacal simpletons,then they will. It is much easier to dismiss you than to consider changingtheir way of thinking or studying a new approach. Why should they waste timeinvestigating something that "obviously can't" scale up or map toreality?
Being "despised and rejected of men" may produce a kind of group pride orfeeling of belongingness, but is not very effective for propagating yourmnemes. The history of science deals with the mnemes that spread widely enoughand survived long enough to make an impression. I have no doubt that university(and monastery) archives are littered with seminal dissertations that neveragain saw the light of day.
Keep up the good work
Paul
Date: Fri, 5 Aug 1994 17:56:35 CST
Subject: Re: B:CP Reactions - Cut 2
Tom Bourbon [940805.1719] >[Paul George 940804 1600]
> {Directed mostly to Bill Powers, but comments welcome fromanywhere}
How can I refuse an invitation like that? Time is short today, so I willonly reply briefly to one or two points.
> Having finished B:CP, I think it is a good piece of work. Too bad somuch of it seems to be ignored in most of the research and the discussionshere, which seem focused on the 'worm's eye view' of PCT (1st & 2nd ordercontrol). Perhaps because the other concepts are deemed 'uninteresting'?:-)
Paul, it is nice to see that you are reading something about PCT. That is asignificant step from the time when you appeared on this net, all filled withthunder and lightning, but with no history of having read any of the material.:-)
Like Rick Marken (940508.0945), I think you understate the case. I wouldplace Powers's B:CP, and his theory of behavior in the same league as Newton's"Principia," or, as I said in the "foreword" to the second volume of Bill's_Living Control Systems_, I think of B:CP in the same class as Aristotle'sideas about "final cause" and William James's ideas about purpose and intention-- only Bill's work is better -- he turned *ideas* about intention into a*science* of intention. But, then, I am a biased lover of only the lowestlevels of behavior and perception. ;-)
. . .
> A large part of my initial impressions of weakness or oversimplicityin HPCT was because you (particularly Rick & Tom) appeared to be sayingmore elaborate structures were unneeded.
Ah, but that was the impression you formed back when you had read nothing,but burst on the scene telling us were using a model that was (or that mightbe) overly simple. Didn't you want us to reply to you on that topic? Should wehave ignored you, or assented without protest to your claims?
> I recognize that most PCT research is focused on levels 1-3 of theperceptual control hierarchy, partially due to amenability to modeling, butcreates the impression that it is all HPCT consists of.
Sorry, but have we read the same book? And which net are you reading from?What you say about the levels we focus on is a pretty serious distortion of thetruth. You have repeated that assertion many times, even though it is patentlyfalse. Why do you do that? I don't understand. Can you clue me in? Remember,though, I probably won't understand your answer if it is more complex than asecond- or third-level explanation. ;-))
> The simple perceptual control loop does not explain all of what iscommonly called behavior, much less psychology.
I'm glad you agree with us on that!
> The reorganization control hierarchy, memory,and the perceptual &memory switches are necessary constructs to deal with common concepts ofplanning, imagination, etc.
The PCT and HPCT models aren't intended to "deal with concepts." They areintended to explain the phenomenon of control.
> Further the levels of the PC hierarchy above 3 are rarely explained,other than by occasional asides - e.g. "The lower six levels are concernedwith control of intensities, sensations, configurations, transitions, events,and relationships. "
Hmm. That' interesting. Are we reading different nets, again? Even thesimplest tracking task reaches the sixth level, at a minimum. And I have shownhow easily a tracking task can be modeled as requiring a program level --that's about the eighth level in a ten or eleven level hierarchy. Are wetalking about the same model, Paul?
> It is not intuitively obvious that these constructs derive from thesimple concept of control. The terms are also subject to variableinterpretation.
Yes. That's why we keep telling you the theory and model are not about thewords. The only way we know for a person to "get it" at the gut level -- to getit "in your bones" -- is to play with the model. Have you run any of thedemonstrations?
> As the [From Bill Powers (940803.1510 MDT)] exchange with comp.aiindicates, your critics and the uninitiated dismiss you because you do notpublicize the 'non-core' ideas, except through casual asides (in myexperience), and so they presume such concepts are not incorporated inPCT.
"In my experience." That is an important qualifier, Paul; your experienceto now is very limited -- not a criticism, but a restatement of what you havetold us. Keep reading -- and run some demonstrations. :-))
While you read, please remember that until *very* recently, only threepeople were working (most often on their own time and at home) to seriouslytest the PCT model in research and simulations. Only three, and I was one. *Nowonder* we have accomplished so little of what needs to be done!
Later, Tom
Date: Fri, 5 Aug 1994 18:14:20 CST
Subject: Re: Appealing to Complexity Worshippers
Tom Bourbon [940805.1801]
>[Paul George 940805 16:00]
>[Rick Marken (940508.0945)], also Avery Andrews (940805.1347)
. . .
> However, if propagation of truth or knowledge is your goal, ratherthan a cloistered purity of thought and purpose, you need to communicate withthe unenlightened.
I'll let that one pass. It would be too easy for me to slip into thinkingof it as an insult, and I don't believe that is what you intended.
> IMHO you need to first communicate that the theory can encompass thecomplexities that many psychologists and commoners ;-) observe in humanbehavior, and then focus them towards the basics. Then they can be amazed athow much so called complex behavior simple hierarchical control can model.But first you have to catch the mule's attention.
Paul, why don't *you* take on that role? Seriously. Make a good-faithattempt to publish some genuine material on PCT applied to the complex issuesof the day in psychology. Don't submit the watered down and distorted versionsso many others have published. You would do a great service for "thecause."
Actually, I believe the only way to get through to the academics in thebehavioral sciences is for us to demonstrate that we can help someone answer avery simple, but very important question, or solve a real problem, then tellthe academics that if they want to do the same thing, here is how it is done --PCT. Nothing else will work. We are trying that approach now. (Bill P., I'llfill you in on latest developments here at the med school a little later --maybe Monday.)
> If you make it easy for them to classify you as monomaniacalsimpletons, then they will. It is much easier to dismiss you than to considerchanging their way of thinking or studying a new approach. Why should theywaste time investigating something that "obviously can't" scale up or map toreality?
But, Paul, even when we have tried to "dumb down" the theory, they haverejected it. There is more to the problem than has yet met your eye.
> Being "despised and rejected of men" may produce a kind of grouppride or feeling of belongingness, but is not very effective for propagatingyour mnemes. The history of science deals with the mnemes that spread widelyenough and survived long enough to make an impression. I have no doubt thatuniversity (and monastery) archives are littered with seminal dissertationsthat never again saw the light of day.
I love it when people tell us we *elected* to live in this relation to"real" scientists! ;-))
Later, Tom
Date: Sat, 6 Aug 1994 02:26:42 -0400
Subject: Re: B:CP Reactions - Cut 2
<[Bill Leach 940805.23:50)] >[Paul George 940804 1600]
Paul, since no one else seemed to mention this...
An additional reason for what you see (or don't as the case may be) is thatPCT is THEORY and HPCT is HYPOTHESIS.
In a very real sense, there is no argument about the "correct" theory forbehavior because there IS only one THEORY and that is PCT.
For a physicist use a label of hypothesis for most of the proposals in therest of behavioral sciences is a serious affront to the term (indeed even theword "science" is "insulted" in such application).
Thus, the theory of PCT has support of experimental evidence and is in theunique position (for this field) of no been challenged by any experimentalevidence.
HPCT OTOH, has a "great deal" going for it but does not have the same sortof experimental evidence. It is a hypothesis in the same sense as such would befound in other "hard" sciences. Careful consideration of the experimental datasuggests that the hypothesis is sound but does not suggest that this can be theonly answer.
You must also remember that this crowd is rather "steeped" in thetraditions of "pure sciences" and is not particularly "big" on trying to do an"end run" on the "community". Little doubt (in my mind) that a major reason forthis reluctance is the "massive" abuse of scientific principles that are seendaily in the popular media.
It is already "bad enough" as far as what can happen to a scientifictreatise in "peer review" without foregoing that step and just "handing" it tothe media. A media whose commitment truth an objectivity might be considered tobe less than always honorable <cough, gag> (watering down that lastsentence was tough for me to do).
Having said that however, I think that Rick emphasized what is the mostsignificant reason; Until one begins to grasp the significance of what closedloop control means at the simplest level and begins to realize that just thephenomenon of control itself teaches that much of what is viewed as "complexbehavior" is rather no more than the physical consequences of the action of acontrol system.
A difficulty that I see is that decomposition of higher level referencescould conceivable proceed in many different ways. People could (and do) argueabout how these different structure details may vary. However, the evidence forthe phenomenon of control is overwhelming and models that can be constructed ofonly a few loops (or even one) provide remarkable fidelity.
There is not the slightest doubt in anyone's mind that the number ofcontrol loops in each of the models is vastly smaller in number than in any ofthe human subjects. However, the issue is that the models are control systemsand work with such high fidelity that the models can reasonably be used forpredictive functions.
I guess what I am trying to say is that PCT is much like electronicsitself. We do not need to explain exactly how a specific electron works its waythrough semi-conductor lattice to understand the operation of a transistor noris such knowledge required to build a computer.
However, the ability to build such a computer is greatly improved when thephenomenon of amplification (in the transistor) is understood.
Thus in PCT, we don't even know how many "wires" there are much less howthey all route. We don't know if there are a bunch of "math co-processors" "upthere" or not. We don't know what sort of "special function processors" mightexist if any. What is rather well established though, is that once a controlledvariable is identified, it is possible to demonstrate reliably that livingbeings control perception.
It is the phenomenon of control that is the "sore spot" not thedetails.
-bill
Date: Sun, 7 Aug 1994 11:26:48 -0700
Subject: What's to explain?
[From Rick Marken (940807.1130)] Paul George (940805 16:00)
> Of course a major reason for 'slow' progress may have been thataffordable computers with sufficient power only became available in the last5-10 years.
Or it could be because, over the last 20 years only three people have beendoing PCT research, and doing it in their spare time to boot, while otherpeople were leaning over their shoulder saying "yeah, but why don't you guysstudy something really important, like the kind of things those AI and complexsystems people are studying?" Sheeez.
Tom Bourbon (940805.1719) and (940805.1801)--
I agree with everything you say, Tom, only more so.
I think one of the things that Paul George doesn't quite appreciate yet isthat PCT is not an alternative explanation of the "facts" that psychology hasalready "discovered". According to PCT, nearly all the facts of psychology arehogwash; there is almost nothing in psychology (including most cognitivescience, AI, etc) for PCT to explain. When we are asked "how would PCT explainsuch and such" it usually turns out that "such and such" is just words used todescribe an (often non- replicable) statistically significant result (averagedover many people) that seemed "psychologically significant" for a brief, trendytime. PCT doesn't explain random phenomena; it explains control.
Best Rick
Date: Tue, 9 Aug 1994 17:34:28 EDT
Subject: Re: B:CP evaluation, part 2
Paul George 940809 1730]
[From Bill Powers (940805.1150 MDT)] and others
Thanks for the response.
> Well, sheesh, how much can we talk about at once? We talk about thethings we have done the most work with. Actually, if I had responded to yourinitial inquiries with a complete summary of everything that is in the PCTmodel, you would have thought I was some kind of nut.
I think there is a middle ground. If you had posted the final PCT model anda description of the hierarchical layers _I_ would have been much happier. Theintro stuff (monthly post) is good but IMHO doesn't identify the existence of alot of PCT concepts, much less the HPCT hypotheses (ack Bill Leach940805.23:50).
If Dag had posted the counter-response to his response on comp.infosystemsyou would see how just asserting the wonders of PCT is received [e.g. "In avery real sense, there is no argument about the "correct" theory for behaviorbecause there IS only one THEORY and that is PCT."- Bill Leach940805.23:50].
It was to the effect of "you can convince yourself you can controleverything if you like" (I trashed the article without thinking). Theevangelical approach comes across as "Cast aside your illusions and openyourself to the truth of PCT. Through faith and study will come trueenlightenment". (I exaggerate a mite ;-). Facts are less important thanperception (speaking of preaching to the choir)
> I'd love to have an FAQ. How much do they cost, and can I get a usedone cheaper? By the way, what is an FAQ?
Assuming you are serious, it is a Frequently Asked Questions list.Customarily a newsgroup has one which is posted periodically and is availablefor downloading from a FTP site. Similar to the monthly posting.
> ...It might turn out to be a really effective document. "How Ifinally got the idea of PCT," or something. In fact, your suggested outlinesounds pretty good to me.
I might be able to in the future, but I will have to get a new copy of B:CPto avoid misstatement. I had to send mine back to the Cincinnati library lastFriday. However, if the Book exists in softcopy (word processors were rare inthe old days;-), the extraction would be faster for you. I would be happy toedit/embellish.
>Tom Bourbon [940805.1801]
> Paul, why don't *you* take on that role? Seriously. Make agood-faith attempt to publish some genuine material on PCT applied to thecomplex issues of the day in psychology. Don't submit the watered down anddistorted versions so many others have published. You would do a greatservice for "the cause."
First of all because a paper from a BSBA in Management Information Systemswould be unlikely to be well received by a PhD review committee of a psychologyjournal. However, If I get into Case Western's Organizational DevelopmentMasters Program next year, I would like to use PCT in some of my projects andperhaps publish. Any suggestions?
> But, Paul, even when we have tried to "dumb down" the theory, theyhave rejected it. There is more to the problem than has yet met youreye.
Granted, but I don't think 'dumbing down' the theory is the key. The ideais to state that structures are hypothesized to address their concerns, andthat they are at least as well justified as their own. Then you can direct themto the details of control theory and 'the test'. Actually, selling toPsychologists may be harder than to others due to vested interests. Theinternet however allows you to reach a larger audience who can benefit (saystudents). Censorship is far more difficult. But.... you have to make thematerials easily accessible.
> I love it when people tell us we *elected* to live in this relationto "real" scientists! ;-))
Nope, but after being knocked around for a number of years you tend to'control for rejection'. Note Bill P's comment to a new poster last week (?)saying 'we've pretty much given up trying to publish outside our own littlejournal' (paraphrased).
>Tom Bourbon [940805.1719]
> Like Rick Marken (940508.0945), I think you understate the case. Iwould place Powers's B:CP, and his theory of behavior in the same league asNewton's "Principia,"...
I tend towards understatement. However, the ideas are perhaps morerevolutionary to Psychologists than to others such as engineers. Ask me againafter I have re-read the book a couple of times. One pass probably isn'tsufficient to extract all it contains.
> Sorry, but have we read the same book? And which net are you readingfrom? What you say about the levels we focus on is a pretty serious distortionof the truth. You have repeated that assertion many times, even though it ispatently false. Why do you do that? I don't understand. Can you clue mein?
That Is my perception based upon what I see discussed, and my understandingof the levels. There may have been other discussions at other times, but thenet has an extremely limited memory. Consider that you as a group may assume alarge body of common knowledge or understanding that may not be shared by allon the list. Also consider that you may not have successfully communicated yourideas. From discussions of the tracking program, it appears only to have 3hierarchical loops. IMHO they all appeared to be focused on just level 2 &3(1 being simulated). However, I haven't seen the code or detailedpapers.
[From Rick Marken (940807.1130)]
> I think one of the things that Paul George doesn't quite appreciateyet is that PCT is not an alternative explanation of the "facts" thatpsychology has already "discovered". According to PCT, nearly all the factsof psychology are hogwash; there is almost nothing in psychology (includingmost cognitive science, AI, etc) for PCT to explain.
I appreciate it, but _saying_ it initially to a body of psychologist (ormost 'outsiders') guarantees that they will dismiss anything else you have tosay out of hand.It makes you _sound_ like crackpots. How did you all respondwhen I just suggested that the canonical model _possibly_ needed elaboration(In fact it had such an elaboration in B:CP)? When you thought I didn't takeyou seriously?. I'm talking marketing or salesmanship here. As I said above,the facts or truth don't matter, it is perception and assumption. They mustthink there is value in PCT before they will evaluate it. And learning dependsupon recognizing that there is something to be learned.
Later Paul George
Date: Wed, 10 Aug 1994 09:15:23 -0700
Subject: PCT Research
[From Rick Marken (940810.0915)]
Tom Bourbon said:
> I love it when people tell us we *elected* to live in this relationto "real" scientists! ;-))
Paul George (940809 1730) replies:
> Nope, but after being knocked around for a number of years you tendto 'control for rejection'.
This implies that we would treat acceptance of PCT as a disturbance. Whynot try "The Test" and see if it is. But you have to be careful about how youdefine the controlled variable. If someone says "I love PCT and I accepteverything you say about it; I think there can be no doubt that PCT is animportant approach to understanding how people control their behavior", wouldyou consider that a disturbance or a non-disturbance to a person controllingfor perceiving "rejection of PCT"?
Tom Bourbon again:
> I would place Powers's B:CP, and his theory of behavior in the sameleague as Newton's "Principia,"...
Paul:
> the ideas are perhaps more revolutionary to Psychologists than toothers such as engineers.
Unfortunately, this is not the case. Engineers reject PCT as vigorously asdo psychologists. Does it really seem to you that the notion of "control ofperception" is taken for granted in control (or any other kind of) engineering?B:CP explains in detail why "control of perception" is the central fact ofpurposeful behavior - - ie. the behavior of all living systems (and somenon-living ones). The power of this fact is extraordinary; it explains howpurposes can be carried out, why behavior appears to be S-R, selected byconsequences or planned output, why organisms get into intra- and interpersonalconflict, etc. No one (NO ONE) before William T. Powers noticed this fact aboutcontrol or understood its implications. That's why B:CP ranks with thePrincipia -- in my mind, anyway.
Me:
> I think one of the things that Paul George doesn't quite appreciateyet is that PCT is not an alternative explanation of the "facts" thatpsychology has already "discovered". According to PCT, nearly all the factsof psychology are hogwash; there is almost nothing in psychology (includingmost cognitive science, AI, etc) for PCT to explain.
Paul replies:
> I appreciate it, but _saying_ it initially to a body ofpsychologist(or most 'outsiders') guarantees that they will dismiss anythingelse you have to say out of hand.
I agree -- and we never do that. We begin (or should begin) by trying toconvince psychologists that behavior IS control -- and showing why this is thecase. Nevertheless, we still get questions about how PCT explains "such andsuch" or other non-control phenomenon. And, indeed, when we answer thesequestions, we are "dismissed out of hand" -- because, of course, it was thefavorite phenomenon of the person who asked. Any suggestions about how to dealwith this problem?
Best Rick
Date: Wed, 10 Aug 1994 11:10:31 -0600
Subject: Hype
[From Bill Powers (940810.0930 MDT)] Paul George (940809.1730)
You make some good points, again. Obviously we all could have done betterin communicating PCT, had we known what kind of response we would get from themainstream. I was pretty naive: I thought that just describing the model asclearly as I could would be enough!
I strongly agree with you about evangelical overstatements of theaccomplishments of PCT. There are many areas in which we have no actual dataabout the worth of PCT as a useful model of behavior. We're working on that,but I truly wish we could all remember to limit our claims to what we can backup with demonstrations and data. There's nothing wrong with extrapolations, aslong as they're labelled as such. But touting PCT as the One True Faith is acertain turnoff for any intelligent person. I'm especially disturbed by claimsthat PCT is a "proven theory" in areas where it is no such thing. If asked, Iwill disavow such claims.
Best to all,
Bill P.
Date: Wed, 10 Aug 1994 14:58:10 EDT
Subject: Re: PCT Research
Paul George 940810 15:00 >[Rick Marken (940810.0915)]
Me
>> Nope, but after being knocked around for a number of years youtend to 'control for rejection'.
> This implies that we would treat acceptance of PCT as a disturbance.Why not try "The Test" and see if it is.
Nailed by sloppy phrasing again. I intended 'minimization of rejection'.More accurately tending to reduce the pain or discomfort brought about byrejection as people tend to desire approval and recognition. You tend to expectrejection in certain quarters. Eventually you stop banging your head againstthe wall. Consider the image of an adult lion being tied up with a string thatrestrained him in cub-hood.
> Unfortunately, this is not the case. Engineers reject PCT asvigorously as do psychologists. Does it really seem to you that the notion of"control of perception" is taken for granted in control (or any other kind of)engineering?
Well, to me it _is_ (on the other hand I am not renowned for conventionalthought processes). Your universe is what you can perceive as compared to whatyou would like it to be (dodging for the moment whether these are intrinsicvalues or remembered perceptions). Any control you attempt can only be detectedthrough perception. People control for illusion all the time ('image', theVedic concept of 'maya'). An engineer or scientist relies upon measurements andassumptions (or associations) about how various measurements relate. Insoftware engineering we are dealing almost entirely with intellectual artifactsusing symbolic techniques.We are dealing with perceptions and assumptions aboutthe world (a model) or another set of symbols. The rules are what ever we agree(or appear to agree) they are. It seems obvious to me that you try to controlresults, not actions.
PCT did reveal to me a number of twists I hadn't previously considered andan amazingly elegant and powerful mechanism to describe the behavior. Yourexperiments demonstrated that a lot can be done by simple feedback and that alot of 'computation' can be eliminated. To that extent PCT _is_ truly abreakthrough.
> Nevertheless, we still get questions about how PCT explains "such andsuch" or other non-control phenomenon. And, indeed, when we answer thesequestions, we are "dismissed out of hand" -- because, of course, it was thefavorite phenomenon of the person who asked. Any suggestions about how to dealwith this problem?
A tough one. In some cases you may be able to rephrase the phenomenon as acontrol problem as Bill P (?) did with the operant conditioning data for therats. In other case you can wave your hands ( cries arise: Heretic! Heretic!:-) and indicate that it would be handled in the higher levels orreorganization in a mechanism not yet studied. You could even be real sneakyand suggest that their model might be incorporated within PCT and encouragethem to try to formulate the model. The approach of 'we haven't addressed ityet, but show me how your theory/concept can be justified in fact orexperiment'(a.k.a. 'we're no worse than you') can also work, albeit not winningfriends. (of course you have never tried anything like this ;-})
In reality however, it is almost a no-win situation. The person in question(particularly a department head or other professor emeritus) would reject _any_theory not his own, as his identity or reputation is based upon it. But, if youcan get them to acknowledge that PCT as a possible interpretation, thoughinferior to theirs, you may be able to convince the next generation.
There is a theory that ideas in science don't change, the proponents of theold ones just die off eventually and no longer sit on review boards, thesiscommittees or chair departments. "When a respected senior scientist sayssomething can be done, he is almost certainly right. When he says it cannot, heis almost always wrong"- R.A. Heinlien
Back to the bushes, Paul
Date: Wed, 10 Aug 1994 16:55:50 CST
Subject: Re: B:CP evaluation, part 2
Tom Bourbon [940810.1655] >Paul George 940809 1730]
>>Tom Bourbon [940805.1801]
>> Paul, why don't *you* take on that role? Seriously. Make agood-faith attempt to publish some genuine material on PCT applied to thecomplex issues of the day in psychology. Don't submit the watered down anddistorted versions so many others have published. You would do a great servicefor "the cause."
> First of all because a paper from a BSBA in Management InformationSystems would be unlikely to be well received by a PhD review committee of apsychology journal.
Do you *really* think they would treat you any worse than they do the restof us? ;-)) . . .
>> But, Paul, even when we have tried to "dumb down" the theory, theyhave rejected it. There is more to the problem than has yet met youreye.
> Granted, but I don't think 'dumbing down' the theory is the key. Theidea is to state that structures are hypothesized to address their concerns,and that they are at least as well justified as their own.
Oops. That bit about "at least as well justified as their own" would bringdown a nuclear barrage. Believe me, I've lived through a few. I *never* try totell editors and reviewers that I am as good as they are -- no any more. Andany attempt I have made at saying, "I'm trying to address your concerns;really, I am," has met a similar fate -- "how can he (me) presume thatcontinuous data on a computer game could be of any relevance/interest to us?"and things like that. Even saying something like, "Gee, folks, I know thisisn't what you are used to seeing, but I'm offering it in the hope that*someone* might find it useful," brings down the hammer -- or is it the ax?Either way, it hurts.
> Then you can direct them to the details of control theory and 'thetest'.
I (we) rarely get that far. You see, we have this real problem: we aretalking about a phenomenon -- control -- that isn't even recognized by mostbehavioral scientists. Somewhere up front, we must say so, and try to show themwhat we are talking about. Just about there is where all hell breaks loose,long before we can get to theory and models and esoteric things like that.:-(
> Actually, selling to Psychologists may be harder than to others dueto vested interests. The internet however allows you to reach a largeraudience who can benefit (say students). Censorship is far more difficult.But.... you have to make the materials easily accessible.
Hey, here we are, for all the world to see. How much more "accessible" canwe get?
>> I love it when people tell us we *elected* to live in thisrelation to "real" scientists! ;-))
> Nope, but after being knocked around for a number of years you tendto 'control for rejection'. Note Bill P's comment to a new poster last week(?) saying 'we've pretty much given up trying to publish outside our ownlittle journal' (paraphrased).
Hmm. Like Rick, I wonder what it would be like to have a reference fornon-zero perceptions of rejection. How much of it do you think each of usprefers? (It's obvious Rick has a reference for much more than I do.) How wouldyou test for that controlled perception? Accepting a few of our articles shouldsend us into deep bouts of depression and reorganization, shouldn't it? I wishsomeone would test that possibility, just so I could find out if it istrue.
>>Tom Bourbon [940805.1719]
>> Like Rick Marken (940508.0945), I think you understate the case. Iwould place Powers's B:CP, and his theory of behavior in the same league asNewton's "Principia,"...
> I tend towards understatement. However, the ideas are perhaps morerevolutionary to Psychologists than to others such as engineers. Ask me againafter I have re-read the book a couple of times. One pass probably isn'tsufficient to extract all it contains.
OK. Let me know when you finish it again and I'll ask again.
>> Sorry, but have we read the same book? And which net are youreading from? What you say about the levels we focus on is a pretty seriousdistortion of the truth. You have repeated that assertion many times, eventhough it is patently false. Why do you do that? I don't understand. Can youclue me in?
> That Is my perception based upon what I see discussed, and myunderstanding of the levels. There may have been other discussions at othertimes, but the net has an extremely limited memory.
The net has *no* memory, and some subscribers approach that condition.;-))
> Consider that you as a group may assume a large body of commonknowledge or understanding that may not be shared by all on the list. Alsoconsider that you may not have successfully communicated your ideas. Fromdiscussions of the tracking program, it appears only to have 3 hierarchicalloops.
In my programs I model the person with only one loop, but it representscontrol by a system with about a sixth-level reference signal. Don't be misleadby the number of loops in the diagrams in our papers.
> IMHO they all appeared to be focused on just level 2 & 3 (1 beingsimulated). However, I haven't seen the code or detailed papers.
Ah, then may I recommend a few papers?
Time to run, Tom
Date: Thu, 11 Aug 1994 09:51:38 -0600
Subject: Re: acceptance of PCT
[From Bill Powers (940811.0615)] Paul George (940810.1500)
I'm following your conversation about PCT research and acceptance at thesame time I'm working on a paper that compares the inverse- kinematic/dynamicmodel of arm control with a PCT model. It seems to me that our problem goesmuch deeper than that of merely gaining acceptance.
There is something seriously wrong with a science that could accept theidea that the brain achieves organized behavior by computing the joint torquesthat will produce it. To do these computations, the brain requires informationabout the masses and moments of inertia of the arm segments, the properties ofmuscle contraction both static and dynamic, the properties of nonlinear musclesprings, the variations in mechanical advantage as the joint angles change, thephysics of dynamic movements, the trigonometry of spatial relationships, thelocation of targets in visual space, and the initial state of the arm in termsof positions and velocities relative to a possibly moving target. The brain isrequired to do computations involving signs and cosines, multiplications andadditions and divisions and subtractions -- and to do all this in real timewith so much precision that after a double time integration, the final resultis the kind of pointing accuracy we observe in the real system. The body isassumed to be as stable as a rock, the muscles to be immune to fatigue, and theenvironment to be free of unpredictable disturbances.
There is an air of dreamlike unreality about this model. It is assumed thatanything a mathematician can do with pencil and paper and symbols, the braincan do with neural currents (without symbols). It is assumed that all theknowledge that external observers have obtained in 300 years of studying thephysics and geometry of the arm and environment is available in real time tothe neural computers, even those of a monkey or a mouse. Just how thisinformation becomes available is not even considered.
So the question naturally comes to mind, "Why should PCT researchers beinterested in acceptance of PCT by people who could bring themselves to believein such models?" Just what would such people be accepting? Another abstractmathematical scheme with no more justification than any others they havebelieved in? Another form of magic? Another kind of prediction that is truesome of the time, for some systems, under some conditions? Is there theslightest chance that they would grasp PCT and use it as a real theory? Orwould they just use it as another "perspective" on nature, to be adopted or notat one's convenience?
PCT does not explain all behavior of all organisms, just as physics doesnot explain all behavior of all matter. But the failures of explanation are ofa kind different from those found in psychology. They aren't statistical; theyare total. There are phenomena that we simply don't understand, and we know wedon't understand them. There is no point pretending that we do understand,until we actually do. What psychology is missing is a concept of this point ofunderstanding where you are simply backed into a corner, and no matter whatalternatives you may think of you are continually forced back to the same view.Everything else is ruled out. If there is a better explanation, and one knowsthere will always be a better one some day, it is simply not available now. Thetrail, for now, ends here. When we reach such a point of understanding there isno choice but to proceed as if it is true.
The kind of understanding that comes out of psychological theories is thesame kind one could obtain with an understandingness pill, like the sense ofgodlike comprehension that some people seem to get from cocaine. It is adecision to believe rather than to look further.
Psychological theorizing does not stop when there is no other place left togo; it stops when an already weak sense of skepticism about one's own ideas iscompletely suspended. So we have dozens of competing "microtheories," allexisting at once and all accepted as part of psychology, and hardly a onedestined to last more than five years.
Is this the field in which we aspire to gain acceptance?
Best, Bill P.
Date: Thu, 18 Aug 1994 15:23:35 EDT
Subject: Re: Ayn Rand and PCT
[Paul George 940818 15:25]
For those who are interested in testing whether PCT would be accepted bythe followers of Objectivisim there is a newsgroup alt.objectivism wherecurrent advocates hang out. I don't have access to it, but there are occasionalcross posts to sci.econ. They may also be found on various of the libertarianforums.
Date: Fri, 19 Aug 1994 08:06:57 -0700
Subject: Accepting vs Learning PCT
[From Rick Marken (940819.0800)] Paul George (940818 15:25)
> For those who are interested in testing whether PCT would be acceptedby the followers of Objectivisim there is a newsgroup alt.objectivism wherecurrent advocates hang out.
I am no longer interested in whether or not anyone in particular wouldaccept PCT; clearly they would. What I am interested in is whether or notanyone would _learn_ PCT, by which I mean learning how to identify controlledvariables (The Test) and how to produce models that control these variables inthe same way that the organism controls them.
In order to learn PCT one must have at least a basic grounding inmathematics, programming, physical science, neurology, physiology, and controlengineering. I think it also helps to be somewhat familiar with conventionalpsychological science, if only to be able to explain what PCT is _not_.
Right now I estimate that there is probably a 4 order of magnitudedifference between the number of people who have accepted PCT (something on theorder of 10,000; I include in this group Reality Therapists as well as researchpsychologists like Carver, Scheier and Lord) and those who have learned it inthe sense I describe above - - that is, done the research and modelling(something on the order of 7). Because a large proportion of those who haveaccepted PCT have not learned much about what they have accepted, there is agood chance that the PCT they accept is not the PCT reflected in the controlphenomena and models studied by those who are doing the PCT research andmodelling.
All this will change when the Center for the Study of Living ControlSystems becomes a reality. One of the main goals of the Center will be to teachwilling members of the approximately 9,993 who already accept PCT how to studyand model what PCT is actually about -- the purposeful behavior of livingcontrol systems.
Best Rick
Date: Fri, 19 Aug 1994 17:45:20 CST
Subject: Re: Accepting vs Learning PCT
Tom Bourbon [940819.1741] >[Rick Marken (940819.0800)]
>>Paul George (940818 15:25) --
>> For those who are interested in testing whether PCT would beaccepted by the followers of Objectivisim there is a newsgroup alt.objectivismwhere current advocates hang out.
> I am no longer interested in whether or not anyone in particularwould accept PCT; clearly they would. What I am interested in is whether ornot anyone would _learn_ PCT, by which I mean learning how to identifycontrolled variables (The Test) and how to produce models that control thesevariables in the same way that the organism controls them.
Yes. The problem is not merely one of finding people who accept PCT. AsRick said later in his post, if we count people who have accepted one oranother version of control theory as a way of thinking about and talking aboutbehavior, they probably number in the low tens of thousands, at least. Theproblem is finding people who accept control theory who are also willing to gothe next part of the journey and learn how to do PCT science -- testing toidentify variables that people control, modeling control in living systems, andapplying the theory _rigorously_, wherever they can. I can probably count thenumber who have gone that far on my personal (limited) assortment of fingersand toes. . . .
> All this will change when the Center for the Study of Living ControlSystems becomes a reality. One of the main goals of the Center will be toteach willing members of the approximately 9,993 who already accept PCT how tostudy and model what PCT is actually about -- the purposeful behavior ofliving control systems.
Surely among that many people are a few who will drop everything else anddo the necessary work. Nothing less will do, not because PCT is a religion thatrequires all neophytes pass through a monastic experience at The Center priorto full initiation, but because PCT is a demanding science that cannot bemastered by taking short cuts. It is _not_ about slogans, catchy one-linephrases, and clever ways to manipulate other people. It is _not_ about how togain a new "perspective" or "framework" from which to reinterpret (and therebycling to) every old idea you already had about behavior. It _is_ aboutfrequently staring your own ignorance, prior conceptions, and lack of skillssquarely in the face, taking a deep breath, then getting to work, doingwhatever _is_ necessary and throwing away whatever is _not_ necessary.
I have joined Rick in giving up on the idea that we should spend any moreof our time trying to get more people "interested" in PCT. Another radical isborn.
Later, Tom
_