Paper presented By William T. Powers at the CSG conference in Durango, July1993.
This paper evolved in the spring of 1993 as a collaborative effort by CSGnetters to identify misunderstandings and "myths" by "devils advocates" amongpeople in academe who have misinterpreted the fundamentals of Control Theoryand rejected manuscripts submitted for publication in scientific andpsychological journals. What follows is version 4 in this effort, crafted byBill Powers. The supporting collection of misleading quotations in theliterature is large. See file DEVIL'S.BIB.
THE DISPUTE OVER CONTROL THEORY
William T. Powers and
The Editorial Board of the Control Systems Group
Feedback control has been seen as a central concept in the behavioralsciences for five decades. But its actual nature has been widely misunderstood,and for this reason its potential and significance have been seriouslyunderestimated, especially since the mid-to-late1970s.
This article is intended to set the record straight on the content andclaims of control theory in general, and PCT or perceptual control theory inparticular, in the context of the behavioral sciences. While the particularadaptation of control theory to be described here has attracted many supportersover its 40-yearhistory, it still represents a minority position. A great many readers ofrefereed journals know it primarily through the way it is represented bycritics --or in some cases by would-besupporters --who have grasped its basic approach to understanding behavior less thancompletely.
Perceptual control theory stands in a peculiar relationship both totraditional disciplines in the behavioral sciences and to branches ofengineering which have explored applications of engineering control theory tothe behavior of organisms. Control theory itself, under any label, offers a wayof analyzing behavior that has been unknown in the behavioral sciences duringmost of their history. Perceptual control theory involves some rearrangementsand reinterpretations of the engineering models, specifically designed tofacilitate the application of control theory to living systems. The result isthat behavioral scientists and engineers alike find objections to PCT, but oftotally different kinds. Objections from behavioral scientists focus ondepartures from traditional ways of interpreting behavior, while those fromengineers focus on the unfamiliar ways of representing control systems that arenecessary to match the model to an organism. We will try here to thread our waybetween the shoals on the one side and the rocks on the other, and show why atleast some students of behavior consider PCT to be as valid an approach as anyothers that have been offered.
THE ORGANIZATION OF A CONTROL SYSTEM
The first thing we must do is to correct several wrong impressions createdlong ago by the too-literaladoption by psychologists of an engineering diagram of a typical controlsystem. Fig. 1 is taken from (__________), but has appeared in many otherpublications going all the way back to the cybernetician Norbert Wiener (1948,p. 132, Fig. 5) and the engineering psychologist ________.
error
Input ---->comparator ---->forward function ----->Output
+ -| |
| |
----<---feedback function <---
FIGURE 1
When a behavioral scientist sees the terms input and output, the naturaltranslation is into _sensory_ input and _behavioral_ output. The above diagramthus seems to suggest that a control system is basically a stimulus-organism-responsedevice, with a feedback loop added. Overall, it converts inputs into outputs inthe normal cause-effectway.
However, in the engineering diagram the input is _not_ a sensory input. Itis a _reference_ input, the means by which the user of a control system can setthe desired value of the output. Feedback action brings the output, or asensory representation of it, to a match with the setting of the referenceinput. The actual sensors are not even shown in this diagram; they are themeans by which the output of the system is sensed and converted into aninternal signal, the one entering the feedback function.
This diagram is also misleading as to the meaning of "output." In anartificial control system, the output is the variable that the customer wantsto be controlled. But the state of this so-calledoutput is not, in general, an immediate consequence of the effector action ofthe system. In a household temperature control system, the controlled variableis not the heat energy being output by the furnace into the ducts, but thetemperature of a room somewhere far from the furnace. In the exposure controlsystem of a camera, the controlled variable is not the opening of the irisdiaphragm or the shaft speed of the actuating motor, but the amount of lightfalling on the photocell and the film. The physical output of a controlsystem's effector is seldom identical to the quantity that is under control.Instead, the effector output is linked to the controlled variable through somephysical process which may be quite indirect and involve changes from one kindof physical variable (iris area, energy output) to another (light intensity,temperature).
Therefore what the engineer calls the output of a control system is only inspecial cases the same as the physical output of the system's effector. Theengineer simply moves the definition of the output to the position of thecontrolled variable, and measures it in units appropriate to the controlledvariable. This, however, is not at all apparent from Fig. 1.
Any non-engineerlooking at Fig. 1 would assume that the output represents the physical, visiblebehavior of the system --in an organism, the patterns of motor output that have effects in theenvironment. The assumption would be that these physical outputs are undercontrol. Working under that assumption, and the previous one that the "input"on the left is a sensory input, a behavioral scientist might well wonder why somuch is made of the feedback connection. At most, such a connection can onlymodify the stimulus-responselaw governing the effects of stimulus inputs on behavioral outputs.
Looking at just the output part of Fig. 1, let us add some details, shownin Fig. 2.
Forward function -->qo----->[fe]------>qc
(effector) |
|
|
<---[fi]<-------------------------------
(sensor)
FIGURE 2
The forward function is, or contains, the effector of the control system.The immediate output of the effector is the state of some physical quantitysuch as a force or a torque. This is labelled qo, for output quantity.
The output quantity is linked to the controlled variable qc through someenvironmental path in which various physical laws come into play. This path isindicated as an environmental function fe.
Finally, the controlled quantity qc affects a sensor, or more generally aninput function fi (which can include a sensor and signal-processingcomputations). This is the feedback function of Fig. 1. The signal emitted bythe input function connects to one input of the comparator of Fig. 1.
Fig. 2 is identical with the output part of Fig. 1 except for some detailsthat remove ambiguities. Instead of showing the feedback path as startingvaguely in the vicinity of the "output" line, a specific physical variable isnamed as the controlled quantity and the feedback path represents its effectson a sensor. The "output" line is broken down into an effector output, qo, anda place to put representations of physical laws, fe, that connect that outputto the controlled quantity.
Functionally, Fig. 2 indicates the same control system that Fig. 1indicates. An engineer looking at the two diagrams would see the second assimply making explicit some details he or she normally takes for granted. Butto many behavioral scientists, Fig. 2 will bring in some new considerations. Wecan make those new considerations even more explicit by completing andrearranging Fig. 2, as shown in Fig. 3.
|sr ref signal
- +|
perceptual sig sp ----->[Comp]--->--se error sig
| |
| sensor effector | system
:::::::::::::::::::[fi]::::::::::::::::::::[fo]:::::::::::::
| | environment
| |
controlled qc <----[fe]<--------qo effector
quantity | output
|
[fd]
|
d disturbing quantity
FIGURE 3
This figure is organized exactly like Fig. 2. It is simply rearranged. Itis actually just like Fig 1. with details added. The plane of separationbetween system and environment, however, is not the one suggested by the firstdiagram. To locate it in the first diagram, one would have to draw a verticalline through the forward and feedback functions, with the environment on theright and the active control system on the left.
This distinction means little in engineering, but in PCT it is essentialfor getting the correspondences between the engineering diagram and thephysical organism right. In Fig. 3, the horizontal line separates the nervoussystem of the organism from all that is not nervous system. Sensors andeffectors lie in the boundary. Notice that in Fig. 3, there is no chance ofmistaking the reference input for a sensory input. The reference signal comesfrom higher inside the behaving system. The sensory inputs are strictlyassociated with the feedback path through the environment. In living controlsystems, unlike artificial ones, the reference signals are not accessible fromoutside the behaving system. The "user" of this system who specifies thedesired level of the controlled variable by adjusting the reference signal isnot an external person, but a higher system internal to the organism.
If a behavioral scientist were looking for a place in these diagrams tointroduce a stimulus, where would it be placed? Working from Fig. 1, theappropriate place would seem to be at the location labelled "input." But asseen in Fig. 3, that location can't be affected from outside the system, atleast not directly. The stimulus input obviously must enter where the sensorsare in Fig. 3. However, in Fig. 3 the sensors are already detecting the stateof the controlled quantity, which in turn is already being influenced by theoutput quantity generated by the effector. The most the stimulus can do is adda contribution to the state of the sensor, which, from a point of view insidethe system, is the same as influencing the controlled quantity. So the stimulusinput becomes a disturbance of the controlled quantity. The external cause ofthe stimulus may or may not actually be sensed. As drawn in Fig. 3, it is notsensed: only the net result of the disturbance and effects of the system'seffector output is sensed. This is the most common case.
The controlled quantity corresponds to a proximal stimulus, a physicalvariable in the immediate vicinity of the sensor. The disturbance is like adistal stimulus, a change in the environment that affects the system onlythrough its influence on proximal stimuli. But these remote causes are not theonly influences on the proximal stimulus: the physical action of the systemitself also affects the proximal stimulus, the controlled quantity. This is thesituation that holds for any control system. The controlled quantity, directlysensed by the system, is affected equally by remote events and by the actionsof the control system itself. And the actions of the control system clearlydepend, at least in part, on the state of the controlled quantity.
This is the famous "closed loop." In the relationship between the activesystem and its environment, there is a little vortex of causality without abeginning or an end. External influences from the environment do not affectjust the sensory input to the system; they affect the causal loop. Thereference signal coming from above does not just produce output; it enters thesame causal loop in a different place, its effects adding to those from theinput function.
It should be clear by now that a control system, properly diagrammed, is anew kind of organization as far as the behavioral sciences are concerned. Theunit of organization is not simply a link connecting higher systems tobehavioral outputs, as assumed in the cognitive sciences; it is not simply alink connecting sensory inputs to behavioral outputs, as assumed inbehavioristic or other empirical input-outputapproaches. It is not any normal kind of causal process in which one can pick astarting point and follow a chain of events through to an ending point. Theclosed loop of causation gets in the way of any conventional kind of attempt totrace cause and effect.
Control theory is the body of mathematical methods that permits us toanalyze and understand the behavior of closed-looporganizations like those in Fig. 3. It is _not_ the theory that "organisms areservomechanisms." Any time that the relationships in Fig. 3 are found,anywhere, control theory is necessary to understand what is happening and topredict the behavior of the system-environmentrelationship.
_Perceptual_ control theory, PCT, is the version of engineering controltheory based specifically on the organization of Fig. 3. In that figure, it canbe seen that a control system controls its own input, not its output. And indoing this, it makes a perceptual representation of that input --a perception, for short --matchan internally-givenreference signal that specifies the desired state of that perception.
This interpretation of behavior is not like any conventional one. Onceunderstood, it seems to match the phenomena of behavior in an effortless way.Before the match can be seen, however, certain phenomena must be recognized. Asis true for all theories, phenomena are shaped by theories as much as theoriesare shaped by phenomena.
THE PHENOMENON OF CONTROL
Traditional scientific approaches to understanding behavior have recognizedtwo kinds of behavior which we will call stimulus-drivenand command-driven.Stimulus-drivenbehavior we can define as changes in behavior that follow from changes in thesurrounding environment. Command-drivenbehavior is a more mentalistic concept, in which at least certain kinds ofcomplex behavior are commanded by internal processes not strictly orimmediately dependent on the environment. Control theory, as we shall see,offers a third choice.
According to the concepts of stimulus-drivenbehavior, the actions of organisms are to be explained in terms of visibleoccurrences in their environments. Under a reflexological theory, stimuli areliterally stimulations of sensory organs, with the resulting neural impulsesbeing routed to the muscles to make them contract; complex behavior resultsfrom complex sets of stimuli that excite many reflexive responses at the sametime. This was the initial concept that psychology inherited from biology andneurology.
Early behaviorism took a somewhat more global view, in which the basicreflexive picture was assumed, but the approach was more abstract. It was notpractical to trace every effective stimulus to the specific stimulations ofnerve-endingsthat it produced, so stimuli came to be defined as stimulus objects or eventsobserved at some distance from the organism. Given the underlying reflexology,it could be assumed that if responses were observed when distal stimulusobjects were manipulated, the appropriate proximal stimulus effects must haveoccurred too. At the response end, it was also impractical to get immersed indetails. Rather than describing behavior in terms of individual muscletensions, the practical approach required looking at larger consequences ofthose muscle tensions, the typical patterns they produced. If some such patternwere observed, it could be assumed to have followed from muscle tensions insome regular way. This more global concept of stimulus and response has had along life.
Even this more global approach was too restrictive for many behavioralscientists. Human beings did not just react to simple objects and events byproducing simple responses; they could react to quite abstract aspects of theenvironment such as complex situations, social influences, and events far inthe past. Such environmental situations could create not only conditionedresponses to current events, but longer-termways of responding which could show up as habits, attitudes, preferences,complexes, conflicts, traits, tendencies, preferences, and biases. Such effectswere seen not in specific acts but as overall patterns of action. This kind ofconcept has also had a long life.
All these versions of stimulus-drivenbehavior account for the bulk of what has been and still is being publishedabout human nature. While proponents of the various versions may contestvigorously and sometimes bitterly with one another, they are all in agreementthat the scientific way to study behavior is to trace out the ways in which theenvironment shapes and directs it. They disagree mainly about what aspects ofthe environment and of behavior are meaningful.
The phenomenon addressed by all these versions of stimulus-drivenbehavior is the observation that behavior is often closely and conditionallyrelated to events in the environment, as effects are related to causes in theworlds of physics and chemistry.
The command-drivenapproach has less scientifically-reputableroots. In this area, the interest has always been in cognitive processes, forinstance: consciousness, language, emotion, intelligence, logic, insight, andgoal-seekingor purposiveness. While many proponents of this view have paid lip-serviceto the more "scientific" ideas across the aisle, the emphasis has not been onenvironmental causation but on the person as a conscious, interpreting,knowing, active agent. The phenomenon addressed in this approach is the factthat many behaviors seem uncaused; even if one calls them "responses," there isno obviously corresponding "stimulus" to account for them. The appearance isthat there is organized activity in the brain that can command behaviorsindependently of immediately antecedent inputs from the environment.
The computer revolution rescued this branch of theory from being orphanedfrom science by offering a link between phenomena of cognition andconsciousness and the workings of computer models of brain functions. A way wasfound to handle "mental" phenomena in physical terms. Many psychologists havechosen not to take advantage of this potential link, regarding thepsychological world as permanently separated from the physical one, butcognitive science as a whole now has at least the potential of working in thesame universe as the rest of science.
Whatever the views of a particular cognitive scientist, there is oneassumption held in common if the brain is thought to be involved at all: thatcomplex brain activities are translated into action through commandsoriginating high in the brain and being elaborated, step by step, until theybecome (primarily) commands for the tensing of muscles. Hence theclassification "command-driven."Those patterned outputs produce, through ordinary physical laws, the globalpatterns we recognize as behaviors. On the output side, at least, the conceptof command-drivenbehavior is in agreement with the concept of stimulus-driven behavior. Where they disagree most directly is in accounting for theimmediate causes of behavior.
This somewhat superficial summary of two main branches of behavioral theoryis not intended to give deep insights into either of them, but to set the stageupon which control theory appeared in the 1930s. The phenomenon of behaviorthat the new control engineers of the 1930s chose to investigate and model doesnot fit comfortably into either of these branches, although both branches haveoften laid claim to it. It is the phenomenon we see when a person controlssomething.
What the new control engineers saw people doing can be seen through anexample. A person operates some piece of equipment, or just that person's ownarms and legs, in a way that affects some variable aspect of the environment.This aspect is also affected by other influences, so the outcome is affectedmore or less equally by the person and by independent forces in theenvironment. The reason for the person's action is that the person wants to, orhas been ordered to, maintain that aspect of the environment in some particularcondition. "I want you to keep an eye on that gauge," the boiler attendant istold, "and adjust the burner to keep that steam pressure nailed at 300 poundsper square inch." Before control theory, there was no theory of behavior thatcould correctly explain how carrying out such a task is possible. This, ofcourse, did not prevent explanations from being offered.
Explaining the phenomenon
From the standpoint of command-drivenbehavioral theory, the attendant's behavior is explained easily: the personadopts the goal of maintaining the gauge at 300 pounds per square inch,constructs a plan for varying the output action on the burner control, andcarries it out. There seems to be little to explain except why the person ismotivated to act in this way. How the task is actually carried out is a matterfor physiologists to explain.
The answer is almost as easy from the stimulus-drivenpoint of view. The person is conditioned to react to fluctuations of the gaugeby moving the hand on the lever or knob that changes the burner heat output.Or: The person is reinforced by receiving a paycheck for responding to thediscriminative stimulus indicated by the supervisor's verbal order in a waythat keeps the gauge at a steady reading, that steady reading eventuallybecoming a secondary reinforcer. Or: The person is influenced by socialpressure to do a good job and please his superiors. Or perhaps: the person is amember of an oppressed class forced by the capitalistic system to bow to theorders of others and engage in this demeaning task. There are many ways toexplain behavior by pointing to something going on in the environment.
None of these answers would have helped the early control engineers,because they were trying to build a device that could do what the person wasobserved to be doing. The engineers didn't need to be told that the person wasactually doing the task. They didn't want to know about antecedent causes ormotivations or social influences. They just wanted to understand how anyonecould do what that person was observed to be doing, whatever the reason fordoing it and whatever the person's or society's attitude toward doing it. As itturned out, nobody had ever before figured out how a person could actually dosuch a task, although some, like William James and John Dewey, had guessedroughly what the right answer would be.
There were three puzzles to be solved (as we can now see theproblem).
First, the causes of disturbances of the variable to be controlled wereinvisible. The steam pressure shown on the gauge would fall if somewhere inanother room or building someone turned on a machine that used steam pressure,or if the energy content of the fuel dropped, or if the line voltage droppedand slowed the furnace's blower, or for any number of other mysterious andunknown reasons. The invisibility of the causes of disturbances made nodifference in the attendant's ability to keep the steam pressure at therequested level.
Second, the steam pressure was indicated on the gauge, but becausedifferent steam pressures might be required at different times, there was noindication of the right steam pressure. All the gauge did was report the actualpressure; its needle did not also indicate the pressure to be maintained. Theonly input corresponding to the right pressure was in the initial instructions,which occurred only once. From then on, no input corresponded to the rightsteam pressure. And there was no input to specify what actions would result inthis end.
Third, the steam pressure, although it was certainly sensed by theattendant, was also in part caused by the attendant's actions. A chain ofcausation could be traced from the gauge reading, into the attendant's eyes,through some hypothetical connections in the attendant's brain (which theengineers had to simulate), through the attendant's muscles, to the burnercontrol, and back to the pressure reading on the gauge. The chain formed aclosed loop.
This third part of the puzzle was the critical part. How could such aclosed loop of causation be analyzed? The gauge reading, in order to be calleda stimulus, had to be independent of behavior. To be called a response, it hadto be dependent on the motor outputs of the attendant. It could not, accordingto traditional thought, be both at the same time.
The stimulus-drivenapproach necessarily would have to find a way to separate stimulus fromresponse. In similar situations, this was usually done by separating them intime. It would be assumed that first a fluctuation of the gauge reading occurs,which leads to a series of events that results in a movement of the attendant'shand and a change in the heat output of the burner. That change alters thesteam pressure in a way that, one hopes, is opposed to the originalfluctuation. Then the cycle can begin again. This at least sounds like aplausible analysis.
The command-drivenexplanation would have an even harder time with this closed loop of causation.In order to formulate a command that would oppose the fluctuation in the gaugereading, a cognitive system would have to know about all potential causes ofdisturbances. But none of the multiple causes of pressure fluctations can besensed. There is no one command that can be sent to the muscles that willresult in opposing an unpredictable fluctuation of the gauge reading. A top-downcommand-driven system can't handle this situation at all.
The control engineers used neither type of analysis. Instead, usingtechniques well-knownin their underlying disciplines, they first characterized each subprocess witha descriptive input-outputequation, and then _they solved all the equations as a simultaneous set_.Sequential cause and effect never entered the picture. Neither plans norcommands were involved. The rule was simply that each variable could have onlyone value at a time. No matter how one variable depended on others, whetherdelays or integrative lags were involved, whether there was amplification inone part of the loop and losses in another, whether static or dynamicrelationships were involved, all the variables in the system had to satisfy allthe equations that pertained to them at a single instant --and at every instant.
That approach, in a nutshell, is control theory.
Comparing the three models
Control theory was expressed by engineers in diagrams like Fig. 1, which,topologically transformed, becomes Fig. 3 in PCT. In Fig. 3 two pathways can beidentified, as indicated in Figs. 4a and 4b.
|sr ref signal
- +|
perceptual sig sp ####->[Comp]##->##se error sig
# #
# sensor effector # system
:::::::::::::::::::[fi]::::::::::::::::::::[fo]:::::::::::::::
# # environment
# #
controlled qc <----[fe]<--------qo effector
quantity # # output
# #
[fd] v
#
d disturbing quantity
FIGURE 4a
#
# sr ref signal
- +v
perceptual sig sp ----->[Comp]##->##se error sig
| #
| sensor effector # system
:::::::::::::::[fi]::::::::::::::::::::[fo]:::::::::::::::
| # environment
| #
controlled qc <----[fe]<--------qo effector
quantity | # output
| #
[fd] v
|
d disturbing quantity
FIGURE 4b
In Fig. 4a, we see a shaded path (#######) starting with the disturbingquantity, going through the controlled quantity at the input, running throughthe perceptual function into the comparator, coming out of the comparator andgoing through the output function, and ending in the output quantity. This isthe pathway equivalent to those envisioned under the various versions ofstimulus-drivenbehavior. The disturbance plays the role of a distal stimulus, the controlledquantity the role of a proximal stimulus, and the output quantity the role ofthe motor behavior (simple or complex) that results. The part of the pathwayinside the system (above the line) corresponds to well-knownneural pathways.
One link in the environment is not on this pathway, the one from the outputquantity, through fe, to the controlled quantity. Recognition of this pathwayis an afterthought in command-driventheories if it appears at all. It appeared in the middle years as the conceptof response chaining, or as reinforcing consequences of actions. The functionfe would correspond to a "contingency of reinforcement" in Skinnerian theory.It was not, however, considered an integral part of behavior: the shaded pathwas the primary path and all other effects were secondary orconsequential.
Fig. 4b shows a shaded path (########) starting with the reference signaldescending from above, then passing through the comparator and the outputfunction to the output quantity. This is the path envisioned in command-driventheories. Somewhere above this diagram lie the cognitive systems that formulateand generate commands that follow the shaded path. A second path is addedoriginating somewhere in the environment and rising toward the cognitivesystems; this path represents informational inputs that form the basis forcognitions. That second path would not influence the downgoing path, so is isnot shown following any existing path in Fig. 4b. If we actually drew a box atthe top and labelled it "cognition," the result would look suspiciously likeFig. 4a. In this diagram, too, the feedback path through fe is an afterthought,indicating effects of one commanded action on the input situation leading tothe next command.
Both stimulus-drivenand command-driventheories treat the closed loop of Fig. 3 either by ignoring it or by trying tosplit it into separately operating parts. Control theory encompasses bothpaths, and makes the external feedback connection a concurrent part of theclosed loop.
Under the control-theoreticanalysis, this diagram is seen as one single entity, the closed loop, with oneindependent input in the form of a reference signal coming from above andanother in the form of environmental disturbances impinging on the loop fromoutside. The signal from above does not act on the outputs of the behavingsystem, but on the closed loop as a whole. The disturbances, likewise, act onthe whole closed loop, not on inputs to the system.
Each of the two traditional concepts recognizes one of the two independentvariables and ignores the other. Cognitive theories recognize that commandsignals are generated independently by higher systems, but fail to recognizethat the output is subject to disturbance from outside the system. Hencecognitive theories assume that the command signal produces a correspondingoutcome in a disturbance-freeenvironment. Stimulus-driventheories recognize independent variables in the environment that act on thesystem, but fail to recognize that the resulting behavioral outputs may also beaffected by signals generated independently inside the behaving system. Thusstimulus-drivenmodels assume that the output is determined solely by the stimulus input. Also,they do not consider the possibility of independent disturbances actingdirectly on the behavioral outcome; disturbances are identified as inputs tothe sensory interface, and the outputs are the sole determinants of observablebehavior.
Neither approach reveals the external feedback path as creating effectsconcurrent with both independent variables, effects that greatly modify theirassumed effects.
We can now see that the PCT model acts as a synthesis of the concepts ofcommand-drivenand stimulus-drivenbehavior, showing how each one corresponds to one aspect of a control system.The synthesis, as often happens, shows both what is right and what is wrongwith the older ideas, and brings out new considerations never covered by eitherone. The situation is similar to that of the blind men and the elephant(Marken, 1992).
HPCT
The diagram of Fig. 3 represents a unit of behavioral organization; it isnot the whole behaving system. This unit can be duplicated and organized intolevels of control, a hierarchical control model referred to as hierarchicalperceptual control theory, or HPCT.
Some idea of the construction of a hierarchical model can be seen bylooking (in a simplified way) at the organization of motor behavior. The lowestlevel of control in a human being consists of spinal reflexes involved inmuscle force generation. The input functions correspond to sensors embedded inmuscles and tendons, which monitor both muscle length and forces applied acrossjoints. The signals from these sensors travel to the spinal cord where theysynapse with motor cells, the length signals with a positive sign and the forcesignals with a negative sign. The motor cells also receive signals from higherin the nervous system; these correspond to reference signals, and the spinalmotor cells play the role of comparators. The output of the spinal cells is theerror signal, which enters an output function made of the contractile elementsof the muscles. The environmental feedback path is composed of the seriesspring components of the muscle and the mechanical laws that convertcontractions into forces on the tendons. Those forces tend to swing the limbsegments about the joints, and this alters the muscle lengths, producingphysical effects that alter the stimulation of the muscle lengthreceptors.
Muscle receptors are also stimulated by effects of minute muscles in thelength-detectingspindle cells. Thus the length signals are really error signals: they representthe difference between actual muscle length and the length of the spindlemuscles, set by gamma efferent signals from higher systems. The muscle spindleis actually a combined mechanical input function and neuromechanicalcomparator. We can thus discern a two-levelcontrol system, shown in Fig. 5.
[Insert Fig. 5 about here ]
[Footnote: The actual arrangement is somewhat different from the one shownin Fig. 5, but is functionally equivalent to it. In the real system, there areindependent alpha reference signals entering the force control system; thesebecome effective mainly when the limb is physically prevented from moving andthe controlled variable becomes applied force. With the limb free to move, thealpha and gamma reference signals have equivalent functions.
The lowest level, shown as Loop 1, controls sensed force, which isequivalent to torque about a joint. The sensed force is compared with thereference force by the spinal neuron, and the difference operates the muscle.This makes the actual force relatively independent of changes in the responseof the muscle to driving signals, and also independent of other sources offorce variation such as inertial effects from limb motions. To some extent thiscontrol system makes the generated force independent of joint angle and limbmotion.
The second level, Loop 2, detects a mechanical effect on muscle length ofthe forces controlled at the first level. This effect is produced by any limbmotions that the forces create. The mechanical comparator compares that effectwith a reference effect specified by a reference signal, and the resultingerror signal enters the spinal motor neuron _as a reference signal for the force-controlsystem_. Disturbances can enter this second level as mechanical disturbances ofthe limb and as loads carried by the limb. The control action renders thesensed muscle length, and hence joint angle, relatively independent of suchdisturbances.
Because of the way we have drawn the boundary between the behaving neuralsystem and its environment, both control loops pass through the environment.The force-controlsystem is the shortest and simplest loop, the controlled quantity being thephysical force generated in a tendon. The muscle-lengthcontrol system is a larger loop in which an effect of the controlled forceoperates through laws of mechanics and physical dynamics to alter thecontrolled quantity of the higher system, the sensed muscle length (whichcorresponds roughly to joint angle). This sensed muscle length is compared witha reference muscle length; the resulting error signal enters the second-leveloutput function, _which consists of the entire first-levelcontrol system_. The second-levelloop is also closed through the environment, now through a slightly less directpath that brings in more global properties of the physical world. A singlesecond-levelcontrol system may alter reference signals in several first-levelcontrol systems, particularly in those employing opposing muscles. All thosefirst-levelsystems comprise the second-leveloutput function.
Now, skipping a few levels, consider how the muscle-lengthcontrol system is used in visual-motorcoordination, a task like reaching out a finger to touch a target. Thecontrolled quantity is a spatial relationship between finger and target, sensedvisually. The perceptual signal indicates the finger-targetrelative distances in three dimensions. If the goal is to touch the target, thereference signals for these three perceptions will all be set to zero (settingthem to a nonzero magnitude would indicate that some non-zerotarget-fingerdistance is to be brought about and maintained). After the required comparisonstage, the signals representing the three errors are routed to all the relevantsets of two-levelcontrol systems as shown in Fig. 6, with appropriate signs, altering thereference signals for muscle length and thus, through the action of the force-controlsystems, altering the configuration of the limb segments and the position ofthe finger relative to the target. The output of the third-levelsystem thus acts through a complex output function consisting of two levels ofkinesthetic control (with many systems operating in parallel) to bring thevisual perception closer to the reference state in the three specifieddimensions.
So the visual control systems, too, involve a closed loop passing throughthe environment. We can easily extend this layering. For perceived fingerposition we could substitute the perceived position of the tip of a pencil heldin a hand. Varying the position reference signals could result in moving thetip of the pencil to make loops and lines; the varying reference signals couldbe the outputs of higher control systems concerned with forming letters. Thereference signals specifying letters to be perceived could be the outputs ofsystems controlling word-perceptions,sentence-perceptions,perceptions of grammar and syntax, and so on as far as one can find reasonableneed for more layers of control. And in every instance, the control loop wouldpass through the environment, where evidence of its operation can be seen.Disturbances of various kinds would be resisted by systems at the appropriatelevel, through automatic variation of reference signals for lower levelsystems.
This is an essential aspect of the HPCT approach. Any specific HPCT modelis falsifiable, because every layer of the model controls specific variables inthe publicly-observableenvironment, and each type of actually-observedcontrolled variable must show resistance to disturbance and hierarchicalrelations to more global variables that match the same phenomena as predictedby the model.
The ideal of a working multilayer HPCT model that matches many levels ofreal behavioral organization is still only a distant prospect. But theexperimental methodology is clear, and rough fits of this model to behavior areeasy to find. A model of the above-describedoperation of visual-motorpointing behavior has actually been simulated on a computer, includingrealistic muscle properties and physical arm dynamics, and it does indeedbehave quite realistically (Powers, 1992).
The basic organization of the HPCT model has higher systems acting strictlyby varying reference signals for lower systems. It is possible for other inter-levelmodes of control to exist; for example, control through variation ofparameters. Also, provision has to be made for reorganizing processes thataccount for the acquisition of new control systems and modification of oldones. All these subjects are somewhat peripheral to HPCT; they are consideredas subjects for future work, when application of the basic model shows the needfor organizations outside its scope.
Building a hierarchical model that works is clearly a huge undertaking.However, the model of Fig. 3 can be applied to behaviors at arbitrary levels oforganization, without the need to specify the lower-levelsystems employed for output or the sources of reference signals. Tasks can beset up in which reference signals remain reasonably constant, and parameters ofa working model can be adjusted for a fit to real behavior. This sort ofapproach has been used, in an exploratory fashion, in many experiments. Thepredictivity of simulations constructed in this way is extremely high;correlations of modeled to actual behavior higher than 0.99 are common. Whilethe behaviors involved are quite simple, the facts discovered in this way areof very high quality, their probability of truth being high enough --millions to one in favor --for use in deductive arguments. For examples, see (Refs)
PCT experimenters, to keep their spirits up, sometimes like to comparethemselves with Galileo rolling little balls down inclined planes. The resultsmay not be obviously earth-shaking,but they are highly reproducible and highly predictive of natural phenomena.There is no telling what sort of science might arise from accumulating suchsimple but utterly reliable facts.
MISAPPREHENSIONS AND MISSTATEMENTS
The introduction of control theory to the behavioral sciences has beengradual and spotty. In part this has been the result of its proponents' onlygradually coming to understand its full meaning with respect to living systems.But a very important component of the problem has been the difficulty thatconventional behavioral scientists have had in grasping the fundamentals ofcontrol theory and seeing how control theory differs from more establishedinterpretations of behavior.
As indicated in Figs. 4a and 4b, there is enough overlap between theconcept of a control system and more conventional concepts to provide temptingopportunities for assimilating the PCT model into older frameworks. Compoundingthe difficulty has been a surprising tendency for scientists who are normallycareful to know what they are talking about to leap to intuitive conclusionsabout the properties and capabilities of control systems, without first havingbecome personally acquainted with the existing state of the art. In many casesthere is a strong suggestion of defensiveness in the misinterpretations, asnormal sequential or causal analysis is used to show that control systemseither can't work or else work according to conventional principles. Neitherassertion is true. The controversy over control theory in general and PCT inparticular has involved factors not strictly of a scientific nature.
We will deal now with some of the major misinterpretations andmisstatements about control theory that have appeared in the refereedliterature. In citing specific examples, we do not mean to blame specificauthors for the errors; most of these errors originated long ago and have beenpropagated by hearsay, attaining the force of myths elevated to the status offacts. If any criticism is warranted, it is for promulgating statements with anauthoritative air without having verified personally that they are justified.Most of the mistakes we will cite are common and understandable; most beginningstudents of control theory go through the same process of trying to make theprinciples of control fit into the causal world with which they are familiar.But beginning students of control theory do not publish their guesses.
Some of the points to be made below will concern the basics of controltheory. Errors at this level will simply be corrected because they indicate amisunderstanding of the basic idea. Other will concern misrepresentations ofPCT as it applies to living systems. Here there is no assertion that PCT isnecessarily the best or only version of control theory or even the best or onlyexplanation of behavior. The point here is that if arguments against PCT are tobe published, they should deal with what PCT actually says and not with amisrepresentation of it.
[HERE ENDS THE PAPER AS OF JULY 24, 1993. THE REMAINING SECTION WILL BECONSTRUCTED FROM EXAMPLES IN THE LITERATURE CONTRIBUTED BY CSG MEMBERS ANDPARTICIPANTS IN CSGNET.]
---------------------
See file DEVILS.BIB for the posts where such examples in the literaturewere discussed. This topic was also discussed as the first item on the agendaduring the 1993 CSG conference in Durango.