Book information:
MIND READINGS: EXPERIMENTAL STUDIES OF PURPOSE
Experimental Studies of Purpose
Richard S. Marken
Contents, Foreword, and Introduction
CONTENTS
Foreword, William T. Powers vii
Introduction 1
1. Purposeful Behavior
The Nature of Behavior: Control as Fact and Theory 11
2. Mind Reading
Intentional and Accidental Behavior: A Control Theory Analysis 35
Behavior in the First Degree 41
3. The Causal Circle
The Cause of Control Movements in a Tracking Task 61
Closed-Loop Behavior: as Control of Input 67
4. Control of Consequences
Selection of Consequences: Adaptive Behavior from Random Reinforcement 79
Random-Walk Chemotaxis: Trial and Error as a Control Process 87
5. Hierarchical Control
Levels of Intention in Behavior 109
Spreadsheet Analysis of a Hierarchical Control System Model of Behavior 133
6. Coordination
Perceptual Organization of Behavior: A Hierarchical Control Model of Coordinated Action 159
Degrees of Freedom in Behavior 185
7. Applications
Human Factors and Human Nature: Is Psychological Theory Really Necessary? 207
FOREWORD
This is a book that can show a willing psychologist how to do a new kind ofresearch. The theme that runs through all these papers is modeling, theultimate way of finding out what a theory really means. Richard Marken is askilled modeler, as will be seen. But he has a talent that goes beyond puttingideas into the form of working simulations, a talent that can be admired but ishard to imitate. He finds the essence of a problem and an elegantly simple wayto cast it in the form of a demonstration or an experiment.
Sometimes the demonstration is so elegant and so simple that it slips pastthe reader without being noticed (more to the point, it slips past journalreferees without being recognized as a fundamental contribution). Nearly everymodel in these papers, which did make it past the referees, is the sort thatought to convey to a reader a straightforward message: if the phenomenon yousee here really works as this model shows it to work, then a whole segment ofthe scientific literature needs to be deposited in the wastebasket.
That is a message that ought at least to be discussed. To appreciate it,however, one has to understand what Marken's models are doing in those paperswhere models are compared against human behavior. If you pay attention, youwill see that they are all close imitations of a human being doing a task. Ineffect, the model, a computer simulation, is placed into the same experimentalsituation that a human participant experiences, and it behaves by producing thesame kind of action the human being produces. The disturbances that are appliedare just as much a surprise to the simulated person as they are to the realperson.
The computer necessarily senses the situation differently than the way theperson does, and produces its actions by different means. But the organizationof the computer program, which represents a guess as to how the human being isinternally organized, is sufficient to create behaviors that not only"resemble" the human behavior, but quantitatively reproduce it. Thatquantitative reproduction is the main point to which you should pay closeattention.
The plotted behaviors of the simulations are not produced by curve-fitting.They are generated in real time in exactly the way the record of the humanparticipant's behavior is generated, one point at a time. Every point on theseplots is a test of the model. What is reproduced is not just a trend, or anaverage, or some abstract characteristic of the whole data set. Instead, thedetails of unfolding behavior are recreated independently, on the basis of thecontinuous functioning of assumed internal processes. The model can be runafter the real behavior, to match its characteristics to those of the realsystem, or before, to predict new behavior under new conditions.
This, then, is truly the ultimate test of an idea about how any kind ofbehavior is created. The idea is cast in the form of a "generative model," onethat will behave as it must behave according to the organization it has beengiven. Once the model is set up as a network of related processes, thetheoretician is committed: whatever that model then does in the experimentalsituation is the meaning of one's idea about the organization. It then nolonger matters what one thought an organization of that kind would do: it isevident what that kind of organization actually does.
The literature of psychology is full of organizational diagrams. In mostcases, these diagrams represent ideas about how some behaviors are caused byinternal processes or external interactions. But in most cases, you will notfind any test of the model inherent in these diagrams. The diagrams aredescriptions of how their authors think the real systems are organized; theyare not working models. It wouldn't be possible, on the basis of thedefinitions given, to cast the models in the form of devices or simulations,turn them on, and see what they do. It's simply assumed that the depictedfunctions and relationships would do the same things that the real organismdoes, because that's how it seems to the author.
This assumption is almost always wrong. In the first place, to repeat, suchputative models are seldom defined in enough detail to turn them into an actualworking simulation. But even more important, even the simplest diagrams withonly a few boxes and arrows carry implications that their own inventors can'tpossibly forsee. The only way to discover those implications is to construct anexample of the diagram with specific numbers and functional forms to make itwork, turn it on, and watch it run by itself. Any experienced modeler will tellyou that the result will be a surprise--usuallyan unpleasant surprise.
All working models do something. What is hoped, of course, is that theirbehavior will be very close to some specific behavior of a real person. Butbefore that hope can even be tested, the model itself must be specifiedcompletely enough to behave. That behavior shows what the consequences of yourassumptions are. Only when those consequences have been generated and examinedcan any comparison with real behavior be made. This is the step that is missingfrom ordinary behavioral research. Ordinarily, the creation of a conceptualmodel ends the process. In truth, however, it is only the beginning. Only whenthe conceptual model is turned into a working simulation can its behavior becompared against the real behavior. Only through that comparison can thedifferences between the modeled behavior and the real behavior be seen. Andonly when those differences--whichcan be alarmingly large are seen can one go back and refine the model to makeit behave more correctly.
Modeling is therefore an iterative process through which one's assumptionscan be systematically changed until the behavior of the model--theactual behavior, not the imagined behavior--isas close as possible to the real behavior. That is the method of modeling, andthe secret of all the successful physical sciences. Where it has been tried inthe life sciences, it has been equally successful. There is every reason tothink that it should be the central method of any enterprise that lays claim tothe name of science.
Most of the papers you will find here are lessons in the use of the methodof modeling. If you look at them that way, you will learn more than just somefacts about behavioral organization. You will learn how to raise psychologicalresearch to a new level of competence. Even though the method of modeling mighttake longer than turning a statistical crank, the result in the end will be anew experience for most experimental psychologists: models that work with greatprecision, and that teach you things you didn't already know. You will developa new eye for nonsense and obfuscation: once you have created a model thatactually works and matches real behavior, you will see just how little sometheories in psychology have to do with reality.
This little collection of papers will someday be required reading in anycourse in psychology. You are fortunate to be able to begin reading themnow.
William T. Powers
Durango, Colorado
April 1992
INTRODUCTION
The papers collected in this book are the result of a decade of research onthe control-theorymodel of purposeful behavior developed by William T. Powers (1973). I decidedto gather them together in a single volume for several reasons. First, I feltthat a collection of papers describing experimental tests and demonstrations ofcontrol theory would be a useful supplement to existing theoretical (Powers,1989) and textbook (Robertson and Powers, 1990) treatments of the subject. Ialso felt that my published research covered a broad enough range of topics tomake a book like this feasible. Finally, and on a personal note, thepublication of this collection marks the end of an era in which my researchfocused largely on what is wrong with current theories of behavior and thebeginning of an era in which my research will focus almost exclusively on whatis right with control theory.
PURPOSEFUL BEHAVIOR
I used to think that it is a scientist's job to show what is wrong with onetheory before proposing a new one to replace it. But I have learned that thingsare not so simple with control theory. The problem is that control theory isnot really a replacement for any existing theory of behavior. Rather, it is anexplanation of a phenomenon that is not even recognized by current theories ofbehavior--thephenomenon of control. Control is the process of producing consistent resultsin the face of unpredictable disturbances. Control can be as simple as keepingyour car in its lane on a windy day or as complex as keeping your businessprofitable in a shifting economy. The phenomenon of control is more commonlyknown as purpose. The driver has the purpose of keeping the car in its lane;the businessman has the purpose of keeping the business profitable. Controltheory is an attempt to explain this kind of purposeful behavior--thatis, to explain control.
Control theory is a comprehensive, scientifically rigorous, andhumanistically satisfying approach to understanding the behavior of organisms.Nevertheless, it has been almost completely ignored in the field that takes thestudy of behavior as its purview: psychology. I believe this is because manypsychologists (and even some control theorists) don't understand the phenomenonthat control theory is trying to explain. Thus, it is appropriate to begin abook about control theory with a description of the phenomenon of control. Thefirst chapter of this book provides a formal description of behavior as aprocess of control. It is argued that a great deal of what psychologists havebeen calling "behavior" is actually purposeful behavior--inother words, control. The chapter describes objective methods for testingwhether or not any behavior involves control. The goal of these tests is todiscover the controlled (purposeful) results of an organism's actions.Controlled results are also called controlled variables. A controlled variableshould vary, but doesn't; it is kept under control. As a car moves down theroad, its position in its lane should vary considerably, but it doesn't. Theposition of the car is a variable that the driver keeps from varying; it is acontrolled variable. Once a controlled variable has been identified, the theoryof control can be used to explain how it is controlled, and why.
MIND READING
Systems that control the results of their actions are called controlsystems. Thus, all living organisms (and some non-livingartifacts, such as thermostats) are control systems. An important (and,perhaps, surprising) characteristic of control systems (living and non-living)is that they control what they perceive, not what they do. This means that itcan be difficult for an observer to tell what an organism is "doing" just bylooking at its behavior (that is, by looking at the results of its actions).The observer must do tests to determine which results are perceived andcontrolled by the organism. These tests are done by applying disturbances tovariables that might be under control and watching for lack of effect. If avariable is under control, then the effect of a disturbance will be canceled bythe organism's actions. This approach to finding out what an organism is doing(that is, what variables it is controlling) is called the "test for thecontrolled variable."
The papers in Chapter 2 describe the test for the controlled variable indetail and show how it can be used to do something very much like mind reading.The test is a means for detecting the intended results of an organism'sactions. Intentions are in the organism's mind (or brain, if you prefer), sothey cannot be directly observed; but they can be inferred by testing for thevariables that the organism is controlling. The test for the controlledvariable can be viewed as a way to distinguish intended from unintended(accidental) results of an organism's actions. The demonstrations described inChapter 2 show that behavior, from a control-theorypoint of view, is a subjective rather than an objective phenomenon. Controltheory shows that it is impossible to know what an organism is doing withoutknowing what it intends to do; that is, without knowing what is going on in itsmind. The test for the controlled variable is a reliable, model-basedalternative to the "operational definition" approach to defining behavior. Theoperational definition says that behavior is whatever the psychologist says itis; the test for the controlled variable says that behavior is what theorganism intends to do.
THE CAUSAL CIRCLE
Control theory explains how control systems control perceptual variables.The basic functional organization of a control system is a causal loop; whatthe system does affects what it senses, and what it senses affects what itdoes. The behavior of this loop has no beginning and no end; there is acontinuous wheel of causality. This creates a problem for psychologists whohave been trained to look at behavior as part of a causal sequence (the lastpart). When confronted with a clear-cutcase of control (such as the tracking tasks described in Chapter 3), thepsychologist is likely to see the organism's behavior as the end result of achain of cause and effect. This chain begins with stimulus inputs and ends withbehavioral outputs. In a tracking task, for example, where the subject uses ahandle to control the distance between a target and a cursor, the target-cursordistance is seen as a stimulus that causes the subject's behavior (handlemovements). It looks like control (keeping the cursor on the target) can beproduced by this type of cause-effectprocess. In fact, it cannot, but the cause-effectview is deeply ingrained in the way we think about behavior, making itdifficult to imagine that behavior could work any other way.
The papers in Chapter 3 show that control cannot be produced by a cause-effectsequence. The stimulus in a control task is just not smart enough to know howto cause the subject to make exactly those responses that keep a variable undercontrol. The variable that looks like the stimulus in a control task is usuallythe controlled variable itself. This is true in a tracking task where thedistance between target and cursor is under control. In fact, target-cursordistance is both a stimulus (because it influences what the subject does) and aresponse (because it is influenced by what the subject does) at the same time.When there is a circle of cause and effect, the old straight line causal viewof behavior just doesn't work. Straight-linecausality cannot produce control. Only a causal-loopcontrol system can keep a variable in a constant or varying reference state.The reference state is the intended state of the controlled variable. In atracking task, the reference state is usually defined as "cursor on target,"but it could be "cursor 1 cm to the left of the target," or "cursor 10 cm tothe right of the target," or "cursor moving back and forth between each ofthese two points." Whatever the reference state, it is the control systemitself, not the environment, that determines its value. It is in this sensethat a closed-loopcontrol system is an autonomous agent. A control system controls its owninputs; it is not controlled by those inputs.
CONTROL OF CONSEQUENCES
By ignoring the existence of controlled variables, it is possible to seethe behavior of a control system as though it were controlled "from outside,"by environmental variables. One way the environment appears to control behavioris through "selection by consequences." This view of control is associated withreinforcement theories of behavior. These theories are based on the observationthat certain consequences of behavior (reinforcements) influence the occurrenceof the responses that produce them. But responses also influence the occurrenceof the consequences themselves. Thus, when a rat presses a lever for food, thefood influences the occurrence of the lever press, but the lever press alsoinfluences the occurrence of the food. Reinforcement theorists have focused onthe influence of consequences on responses and have concluded thatreinforcements control behavior. Control theorists, on the other hand, takeinto account the influence of responses on consequences and show that it ismore appropriate to view behavior as the control of reinforcement.
The papers in Chapter 4 show that the effects of consequences on behaviordepend on what the subject is trying to do--thatis, what variables the subject is trying to control. The first paper shows thatthe consequences of responding, even if they are considered reinforcing, do notcontrol behavior. When the consequences of behavior are random, subjects arestill able to produce consistent results--thatis, they can control the consequences of their actions. The second paperpresents a model of how subjects control consequences, even when theseconsequences are random. An important feature of the model is its ability tospecify a reference state for the consequences of its actions. The model showsthat what constitutes a reinforcement is determined (and can be changed) by theintentions of the organism; reinforcements themselves have no intention tocontrol behavior.
HIERARCHICAL CONTROL
Organisms control variables in order to control other variables. Theycontrol muscle tensions in order to control pint angles, they control jointangles in order to control body movements, and so on. There seems to be ahierarchy of control systems involved in the production of behavior. Thishierarchy is evident when we look at the behavior of organisms. A birdretrieves twigs in order to build a nest; a composer draws little black dots onpaper in order to create a symphony. The idea that behavior is organized as ahierarchy is not new, but control theory gives it a new slant by suggestingthat behavior involves the control of a hierarchy of inputs, rather thanoutputs. In the control-theoryhierarchy, higher-levelsystems tell lower-levelsystems what to perceive, not what to do. These higher-levelsystems specify the reference states for perceptions that the lower-levelsystems are trying to control. In other words, the higher-levelsystems determine the purposes of the lower-levelsystems. Similarly, the lower-levelsystems are used to achieve the purposes of the higher-levelsystems.
The papers in Chapter 5 give examples of hierarchical control and show howa hierarchical control system works. Hierarchical control can be seen in therelative timing of control actions. In a control hierarchy, lower-levelsystems must operate faster than higher-levelsystems. Higher-levelsystems cannot produce an intended result before the lower-levelsystems have produced the results on which it depends. This nesting of controlactions can be seen in the differential speed of operation of control systemsat different levels of the control hierarchy. Lower-levelsystems not only correct for disturbances faster than higher-levelones; they carry out this correction process during the higher-levelcorrection process. The lower-levelcontrol process is temporally nested within the higher-levelcontrol process. This nesting is evident in the experiments described inChapter 5, where a faster lower-levelsystem controls the distance between a cursor and a target. This system keepsoperating as usual even when, due to a change in the relationship betweenhandle and cursor movement, there is an increase in perceptual error. Normaloperation is restored only after a slower higher-levelsystem has time to control the relationship between handle and cursormovement.
The second paper in Chapter 5 shows how a hierarchy of control systemsactually works. A three-levelhierarchy with four control systems at each level is simulated in an electronicspreadsheet. The spreadsheet model makes it possible to observe the dynamicbehavior of the control systems as they correct for the effects ofenvironmental disturbances and changes in higher-levelreference signals. The systems at each level of the hierarchy control adifferent perceptual aspect of the environment. The spreadsheet model showsthat each system in a control hierarchy is able to achieve its own purposewithout preventing other systems from achieving theirs at the same time.Nevertheless, conflict between control systems can occur. The model shows thatthese conflicts result when two or more systems try to control the same (or asimilar) perceptual variable. The model also shows that conflict is the mostdebilitating thing that can happen to a control hierarchy, short of physicaldestruction of its component parts. Conflict makes it impossible for higher-levelsystems to achieve their purposes. The model provides an understanding of thenature of conflict that should be of interest to theorists andclinicians.
COORDINATION
The hierarchical control model provides an elegant solution to one of themost difficult problems for theories of behavior: the problem of how anorganism coordinates all the actions required to produce an intended result.For example, how does an organism coordinate the temporal pattern of movementsof its arms and legs in order to produce the result called "walking"? Sometheories assume that coordination is the result of complex computations thatcommand just the right actions at just the right time. Others assume thatcoordination is the result of complex physical constraints on how the organismcan move. But none of these theories explain how coordination is possible in aconstantly changing, disturbance-proneenvironment. The hierarchical control model solves this problem by controllingthe perceived consequences of its actions, rather than the actionsthemselves.
The papers in Chapter 6 show how control systems can generate preciselycoordinated actions so as to produce consistent behavioral results. Thiscoordination is achieved in the face of unpredictable and undetectabledisturbances, just as it is by real organisms in the real world. Thecoordination problem is solved automatically by the disturbance resistance thatis characteristic of the individual control systems. As control systems act tocontrol their own inputs, they often disturb inputs controlled by other controlsystems. The actions of these latter systems appear to be coordinated withthose of the former simply because the actions of any control system compensatefor the net effect of disturbances to the variable that it is controlling. Itis not necessary to make detailed computations of outputs in order to producecoordinated behavior. In a control hierarchy, it is only necessary to computethe discrepancy between the intended and actual consequences of those outputs.Coordination flows naturally from the process of controlling perception.
APPLICATIONS
Finally, the proponent of any scientific theory likes to think that thetheory might be of some value to humanity. A good theory helps us understandsome aspect of reality (control theory helps us understand the reality ofpurposeful behavior), but it also helps us solve some of the problems ofdealing with that reality. In the final chapter, I make some suggestions abouthow control theory might help us solve some of the problems of dealing with thehighly technological environment that has grown up around us in the last fewdecades. These problems are the concern of people in my own profession--humanfactors engineering. The human factors engineer tries to design technologiesthat can be used easily, safely, and productively. In order to do this, theengineer must know something about how people use these technologies to achievetheir own purposes. In other words, the engineer must know something about howpeople control. The final paper in this book shows how control theory can helpengineers design a more "user friendly" environment, one in which human controlsystems are best able to do what they must do--control.
Richard Marken 1992