[Martin Taylor 980131 21:50
As the Authors of the several papers are only too well aware, for some
overly long time I have been putting together a special issue of the
International Journal of Human-Computer Studies on Perceptual Control
Theory. They may be pleased to know that I will very shortly be sending
in the papers to the general Editor of the journal, after which I suppose
that they will be scheduled for publication.
Before I send them, however, I would like the authors (in particular), and
any other reader of CSGnet to comment on the following draft of my Editorial
for the issue. It is an attempt to set the stage of the issue for a reader
who either has not heard of PCT or who has dismissed it for some reason.
In light of the many dismissals of PCT, there is a section on the objections
that have been raised against PCT, and I hope that those whose writings have
been subjected to objections I have not listed will let me know, so that
those objections can be included. Any other comments are also wlecomed.
Martin
···
--------------------------------------
Editorial: Perceptual Control Theory and its application
The idea underlying Perceptual Control Theory has been known at least
since Aristotle, and probably much longer. It is that people act so as
to get what they want, in the face of unpredictable events in the world
in which they live. Novelists and playwrights have known this, but apart
from a few shining exceptions such as William James, scientific
psychologists seem to have largely ignored it in favour of theories
based either on the notion that people will tend to behave the same way
when confronted again with the same outer world situation, or on the
"cognitive" notion that people preselect their actions to achieve their
goals, ignoring the problem that preplanned actions will work if and
only if the world is as the plan expects it to be. Neither of these
underlying notions is appropriate to the world of everyday life, to
which Perceptual Control Theory is addressed.
Neither Aristotle nor William James had access to the technical
knowledge required to turn the basic idea into a testable theory. That
knowledge developed in the field of electrical engineering before and
during the Second World War, under the name of servo theory or control
theory. Norbert Wiener introduced the concept of Cybernetics, using the
theory of control systems as a basis for studying "The human use of
human beings." In the early 1950's W.T. Powers noted that "acting to get
what is wanted" is the defining characteristic not only of humans and
other animals, but also of engineered control systems, and began a long
investigation into the implications of this fact, leading in 1973 to the
publication of "Behavior: the Control of Perception," a book that is
still considered the "bible" of Perceptual Control Theory. This issue of
the International Journal of Human-Computer studies consists of papers
solicited both to illustrate the fundamental theory and to exemplify the
wide range over which the theory can be, and is being, applied.
The Necessity of Perceptual Control Theory
There is one overwhelming fact about life: to survive, any organism must
in some way stabilize its essential internal chemistry in the face of
disturbances from a turbulent outer world. Every living thing that we
see is a member of a species that behaves so that at least some members
of the species stabilize their internal chemistry long enough to
propagate their genes.
Protection against the buffeting of the outer world can be done in two
ways, and life employs them both. The first way is to develop passive
armour, such as a membrane, skin, or shell. But the armour cannot be so
perfect as to isolate the organism completely from the outer world; if
it did, the organism would die an entropic death. Just to sustain a
minimal internal organization, any living thing must at least take in
high-quality energy from the outer world, and excrete disorganized waste
energy. Most do much more.
The second way an organism can protect itself against the disturbances
of the world is to counter them as they occur, actively and powerfully.
To do this, the organism must be able to sense important states of the
outer world, it must be able to compare the sensed states with desirable
conditions for those states, and must be able to act so as to bring
about conditions conducive to stabilizing the organism's internal
chemistry.
"To sense" means to alter some internal state, such as a chemical
concentration, or a neural firing rate, in correspondence with changes
in the state of something in the outer world. In Perceptual Control
Theory, such an internal state is called a "perceptual signal," and the
value of a perceptual signal is a "perception." "Perception," in PCT,
carries no connotation of consciousness. It is just the value of a
signal. Not only living things but also inanimate objects such as
computers contain signal values. If those values correspond to states in
the world outside the computer, they may legitimately be called
"perceptions" in the context of Perceptual Control Theory.
Bringing a perception of some physical state to a desired (reference or
goal) value with which it is compared, and maintaining it there, is
control in the strict engineering sense. The perception of the external
state is what is stabilized, not the external state itself - hence
"Perceptual" Control Theory. For this reason, PCT has a core tenet: "All
behaviour is the control of perception."
If a computer acts so as to maintain some signal value corresponding to
a state of the outer world at a reference level, it is controlling its
perception, just as a living organism might do. This does not mean that
the computer is considered to be alive, but it does mean that the
analytic techniques applicable to control can be applied to the relevant
operations of the computer, just as they can to the relevant processes
of a living thing.
Control systems: engineered and human
An engineered control system is designed to set some variable to a
desired value and maintain it there. We will use the aiming direction of
a gun as an example in this editorial. The control system is supplied
with a reference signal, the level of which corresponds to the desired
value of the controlled variable-the desired aiming direction. The
current state of the variable-the actual aiming direction-is not
directly available for comparison with the reference, so before the
comparison can be made, the present aiming direction must be sensed and
transformed into a signal that can be compared with the supplied
reference value. If the two values disagree, the control system produces
output that drives actuators that alter the variable in such a way as to
reduce the disagreement, thus completing a "negative feedback loop."
Whatever the direction and magnitude of the sensed deviation of the gun
aiming direction from its desired value, the negative feedback loop
reduces the error.
Internal to the control system, the variable that is controlled is never
the actual aiming direction of the gun. Always it is the sensed and
transformed signal derived from the aiming direction of the gun-the
"perception" of the aiming direction. It is always perceptions that are
controlled, not the outer world states to which they correspond.
In the case of a fixed land-based gun, a gunner who knew the current
aiming direction could bring the gun to its desired aiming direction by
driving a motor of a known torque for a predetermined time. If, however,
the gunner's knowledge of the motor torque or of the gun's moment of
inertia and frictional resistance was in the slightest degree
inaccurate, the resulting direction would not be what the gunner wanted.
To assure that the gun was pointed correctly, the gunner would still
have to look at the gun's actual position and correct any remaining
error, completing a negative feedback loop through himself.
Neither can the aim be satisfactorily set by driving a motor a
predetermined amount if the gun is on a ship subject to the vagaries of
the waves. But a control system, whether it be engineered or inside a
human gunner, can keep the gun aimed in the right direction despite
variations in motor torque or the effect of the waves, if it continually
senses the current aiming direction and compares the transformed signal
value with its reference value.
The reference value for the aim of the ship's gun is determined by
considerations that have nothing to do with the motions of the ship, and
everything to do with the warlike intentions of the gunner. The gunner
wants to know that the gun is pointing at the intended target,
regardless of the wave action. The gun-aiming control system allows the
gunner to ignore the waves, and to concentrate only on choosing the
right direction in which to aim.
The gunner can choose the right aiming direction only by selecting a
target and using some instrument to determine its direction. If the
target moves, the gunner must determine its direction anew and
communicate the new target direction as a changed reference signal value
to the gun control system. The gunner acts as part of a negative
feedback loop that corrects the error between the aiming direction
appropriate to the chosen target and the aiming direction asked of (and
produced by) the gun control system.
Using radar, an elaborated control system could relieve the gunner not
only of the need to keep observing the gun direction, but also of the
need to keep observing and measuring even the direction of the target. A
second-level control system could be given a reference signal that
identifies the target among the radar returns, allowing it to lock onto
the gunner's desired target among possibly many potential targets. If
the chosen target moves, this second-level system could then send to the
first-level control system a new reference value for the desired aiming
direction, bringing the gun direction back onto the shifted target.
The reference signal for the second level control system is what the
gunner sees as "target identity." An error at the second level would be
seen as the system having locked onto the wrong target, whereas an error
in the first-level control system would be seen as a bad aim at the
correct target. The perceptual signals at the two levels correspond to
characteristically different things in the outer world. This is true
whether control is performed by a mechanical system or by the human
gunner.
There is nothing mystical about perceptual signals in either the
electronic system or the human: whether a perceptual signal corresponds
to the gun aiming direction, with the identity of the target, or with
anything else in the outer world, inside the control system it is just a
value of some quantity such as an electrical voltage or a neural spike
rate. Such a signal has no "meaning" in itself, though it may have a
meaning to an outside observer who can see both the signal and the
environmental phenomenon that determines the signal value. All the
control system does is to keep the electrical voltage or the neural
spike rate near its reference voltage or spike rate. An external
observer sees that the gun aims aims accurately and aims at the right
target. The control systems do not. They "see" only that their
perceptual signals have the values that they should have.
Either the human gunner or an engineered control system can keep the gun
trained on a moving target despite variations in the power of the
gun-motor and the motion of the platform. Either the human gunner or the
engineered control system creates a negative feedback loop that acts on
the environment (the physical gun) so that perceptual signals
representing different aspects of the environment (the aiming direction
of the gun, and the target on which the gun is to be aimed) are brought
to and maintained near reference levels. The control system may be
mechanical, electronic, or inside the human. Functionally, it does not
matter.
According to Perceptual Control Theory, all behaviour is of this kind:
actions are performed to affect the values of perceptual signals so as
to bring them closer to reference values. Elementary control systems
like the gun-aiming control system and the target-following control
system are organized into a layered network or hierarchy, in which the
reference value for any one elementary control system is derived from
the outputs of one or more higher-level control systems. At the highest
level in the hierarchy, there is no reference level as such, but only a
negative feedback loop that stabilizes the perceptual signal at an
arbitrary "reference" value that can be called "zero."
Calibration and Learning: Reorganization
Consider now how perception relates to "reality"-how perception can be
calibrated. What might happen if the sensors that signal the aiming
direction of the gun are miscalibrated? Suppose that when the gun is
actually aiming north, the sensors report it to be aiming
north-northwest. If now the gunner (or the second-level control system)
provides to the gun-aiming control system a reference signal
corresponding to a perception of "North" (where the gun is actually
pointing) there will be a mismatch between the perceived and the
reference direction. Because of the mismatch, the gun will slew round
until the direction sensor reports that it points north, though it
actually points north-northeast. All will seem to be well, since without
using a different sensor there is no way that the gunner or the
second-level control system could tell that the gun is pointing wrongly.
When the gun is fired, however, the fall of shot may be observed using
sensors different in nature from those used to signal the aiming
direction of the gun. If the shot falls to the east of the target, the
observation of the miss permits a correction to how the gun is aimed, in
preparation for the next shot. A new negative feedback loop has come
into play, but one that acts on the existing aim control system, not on
the gun itself. The new correction says "when the sensor signal provides
a perception corresponding to north, treat it as north-northwest." The
aim control system will then equate a desired reference aim of "North"
with a perception derived from a sensor signal corresponding to
north-northwest, and in the real world the gun will be aimed accurately.
In the language of Perceptual Control Theory, the new control system has
"reorganized" the existing two-level control hierarchy by altering the
values of the perceptual signals that correspond to particular states of
the world. This reorganization could be done only as the action of the
two-level hierarchy is tested against reality by observing the fall of
shot.
"Fall of shot" is a perception for which there is no signal in the
gun-aiming hierarchy. It is a signal from a different sensor, one that
relates to the gun's ultimate task of firing a shell that hits the
intended target. But the effect of an error in the "Fall of shot"
perception is a change in the parameters of the gun-aiming perceptual
control hierarchy. In Perceptual Control Theory, this is
"reorganization": a change in the perceptual control hierarchy induced
by error in some perception related to the ultimate task of the
organism-hitting the target, in the case of the gun. Reorganization has
many forms, of which we have illustrated only one-a change in the
function that relates the value of a perceptual signal to the relevant
sensory input.
The reorganization that corrects the gun's miscalibrated direction
sensor has a direct analogy in experiments on humans wearing spectacles
that have prisms for lenses. Many such experiments were conducted
between the 1930's and 1960's (e.g. Kohler, Held, J.G.Taylor). Several
of these experiments compared the relative ability of people to "adapt"
to the prism spectacles when they were able to act on the world as
compared to when they were moved passively around the world. Adaptation
(subjectively, the perception of the world as "normal") occurred much
more quickly when the subjects could act on the world than when they
could not, and occurred with respect to those aspects of the world the
subjects could affect more quickly than for those aspects over which the
subjects had no influence (see particularly JGT and Held). According to
Perceptual Control Theory, these results could hardly have been
otherwise, if the relationship between sensory data and perceptual
signals is affected by a control system that reorganizes according to
the success of the perceptual control system in performing the task for
which it is ultimately responsible.
In the case of the gun, the task for which the perceptual control system
is ultimately responsible is that the gunshot should hit its target.
What is that "ultimate" task in the case of the human? Powers argues
that it is to keep the internal chemistry of the body in a state that
sustains life, at least long enough for many individuals to pass on
their genes to the next generation. The "internal chemistry" is
represented by a set of what Powers calls "intrinsic variables" that
correspond to the chemistry in the same way that perceptual signals
correspond to states of the outer environment. The individual is born
with a set of reference values for these intrinsic variables, and if the
actual, internally sensed, values deviate from their reference values,
the reorganizing control system acts to reduce the deviation. But it
acts by changing the perceptual control hierarchy, not directly on the
outer world, just as the control system that observes the fall of shot
acts on the gun-aiming control hierarchy, not on the gun.
Objections to Perceptual Control Theory
The idea that humans operate as a network of elementary control systems
has been discounted on various grounds, largely specious. The first
objection is the intuitive notion that the theory must be too simplistic
to account for the manifest complexity of behaviour; to this one can
only ask in return whether the simple electromagnetic interactions among
a small variety of simple atoms are too simplistic to account for the
beauty and complexity of the chemical (and biochemical world). Simple
interactions among simple things can very quickly produce results of
astounding complexity.
A second objection, one that sounds more technical, is that control
systems are too slow to account for some observations. To counter this
objection, control theory must be compared with the proposed
alternatives, of which there are two classes: direct stimulu-response
linkage, and preplanned action. According to the notion of direct
stimulus-response linkage, when some stimulus pattern is observed in the
world, "immediately" an appropriate action is performed that acts on
whatever produced the offending stimulus pattern. The objection to
control systems here is that the control action will be too slow, as
compared to the direct stimulus-response linkage. This is a strange
argument, since the connection through the control system is no less
direct than is a stimulus-response connection. Any intrinsic delay in
the effect of the action on the world is the same in the two systems,
since they are subject to the same delays, both inside the person and in
the effect of the person's actions on the outer world. Furthermore, the
control system has a technical advantage in speed, since it inherently
has output amplification, which speeds the action in a way not available
to a directly linked system without feedback. As a demonstration that
control systems are slower than the alternatives, this first alternative
fails spectacularly.
The second class of alternative that is said to be faster than a control
system is a preplanned action system that anticipates the need for
action to produce an intended future state of the world. Such a system
can accommodate delays in the effect of action on the world by starting
the action early enough that when the need arises, the action will have
had its desired effect. Here, the objection is that the control system
cannot act until a changed perception causes an error (deviation from
the reference signal). But this is not true, since a higher-level
control system can alter the reference signal well in advance of the
need for action. To see how, let us return to the gunnery analogy.
The two-level hierarchy, as discussed above, has an upper level control
system that asks a lower level control system to keep the aim of the gun
in a direction corresponding to the changing direction of the target.
But there is a time lag between the moment the shell leaves the gun and
the moment it arrives at the target, by which time the target may have
moved. If the higher-level control system were to ask for the gun to be
aimed at the place where the target would be when the shell landed, the
shell would hit it instead of falling behind it. How could this be
achieved? In the simplest way, the target-following reference value
could be replaced by a value built not just from the current position of
the target on radar, but also from the time derivative of the position.
The resulting output signal that tells the lower control system where to
aim would then be correct if the target maintained its velocity. The
effect would be just as fast as with a preplanned action system, but
would be based on observation of the current and recent behaviour of the
target.
The objection that control systems are too slow fail when compared with
the non-control alternatives. A well designed control systems is
actually faster than the corresponding direct stimulus-response linkage
when dealing with unexpected events in the outer environment, and is no
different from preplanning systems when dealing with a predictable need
for action. A control system has a further advantage over both
alternatives, in that it does not depend strongly on the consistency of
the effect of action.
A third kind of objection to seeing humans as control systems might be
called the "discrete-perception" or "decision-making" objection: humans
are cognitive beings, using logic to make decisions, whereas control
systems are analogue devices that correct errors incrementally. The
objection is that a human chooses one course of action and executes it,
whereas a control system just outputs some signal that is a function of
its present and past error, implying no choice among alternative
actions.
The "decision-making" objection is more subtle than the erroneous
"too-slow" objection. The "discrete-perception" aspect of the objection
is readily dismissed. Both electronic and neural systems have several
mechanisms whereby the value of a perceptual or a reference signal may
be constrained to remain close to one or other of a binary pair of
values that we can call "A" and "B." If the reference signal takes on
the value labelled "A" and the perceptual signal happens to be "B," an
error signal will be generated, leading to output that alters the
reference values for (possibly continuous-valued) lower-level control
systems until the perceptual signal takes on the value "A." Control of
discrete values of a perceptual signal is equivalent to the control of
perceptual categories: I want to perceive a "success" indication on my
computer screen, as opposed to a "syntax-error" indicator.
To consider the "decision-making" part of the objection, we return once
more to the gunnery analogy. Thus far, we have considered two levels of
control: one that keeps the gun aiming in the desired direction despite
the vagaries of platform motion, and one that changes the desired
direction to keep the gun aiming at a desired target as the target
moves.
In order for the second-level control system to operate, we have to
assume that the radar return has been processed in some way so that the
direction and distance to a potential target is continuously available.
This direction, possibly together with its derivative, provides a
reference value for the gun direction, which is compared in the
second-level control system with the perceptual signal corresponding to
the gun direction. We now add the assumption that the processed radar
return embodies not just one direction-distance pair, but a list of
directions and distances to several potential targets. In the example so
far, the human gunner has chosen the target from this set of
possibilities according to a strategic rule. We now add a third level of
control to make the choice (a function that in real life presumably
would not be left to an automated system).
For illustration, we use a trivial strategic rule: aim at the closest
target. To aim at the closest target requires that the closest target be
perceived and selected as the one to be tracked in providing the
reference direction for the second-level control system. To perceive
"the closest target" involves what Powers calls a "logic-level"
perception. At the logic level, perceptions are thought to correspond to
Boolean relationships-AND, OR, XOR, GREATER THAN, and so forth. A
reference perception might be, for example, that the perceived distance
of the chosen target be LESS THAN that of each other potential target.
The AND of all these logical comparisons would then have a reference
value of TRUE. A change of reference value to FALSE directs the gun to
ignore the closest potential target and to aim at any of the other
potential targets, each of which would satisfy the reference value for
the logical variable.
A fourth kind of objection to Perceptual Control Theory is what we might
call the "inverse kinematic" objection. The argument is that the inverse
kinematic equations are underdetermined, because there are so many ways
to accomplish any given desired end result. The consequence is that the
appropriate control actions cannot be uniquely specified for execution.
The "inverse kinematic" objection is invalid when applied to Perceptual
Control Theory, though it does present a problem to any theory of action
that involves precomputing ways to achieve a planned result. In
Perceptual Control Theory, action is not precomputed. A single
elementary control unit merely supplies output when the value of its
perceptual signal does not match the value of its reference signal. This
output may be distributed through an arbitrary number of mechanisms that
might affect the outer environment so as to influence the perceptual
signal. So long as the cumulative influence of this output on the
perceptual signal is such as to reduce the magnitude of the error, the
desired result will be achieved. One of the mottos of Perceptual Control
Theory, along with "All behavior is the control of perception," is "Many
means to the same end; Many ends by the same means." The inverse
kinematic objection turns out to be more of an argument in favour of
than an objection to Perceptual Control Theory.
A more interesting question inspired by the inverse kinematic objection
is: given that a perceptual control system that operates by setting
reference values for lower-level control systems will achieve its
intended objective, will it do so in a way that is like the way a human
does the same thing? For example, will the trajectory of a simulated
hand reaching for a target be like the trajectory of a human hand under
the same circumstances?
It appears that the answer to these questions is a qualitative "Yes." In
this issue, Powers describes a simulation of a system that uses stereo
vision and an arm jointed like a human arm at the shoulder and the elbow
to keep a "fingertip" on or close to a target that moves in a 3-D space.
A current project to follow up this simulation is a 14-degree of freedom
arm that has weights, elasticities, and muscle strengths like those of a
human arm, to test whether the trajectories that look qualitatively like
those of human arms will be quantitatively accurate when following a
moving target.
This editorial is not the place to attempt a full description of
Perceptual Control Theory. The basic introduction in Powers "Behavior:
the control of perception" remains valid for the most part. A selected
bibliography of writings on Perceptual Control Theory is available
through the World Wide Web site of the "Control Systems Group," which is
devoted to Perceptual Control Theory (at http://www.ed.uiuc.edu/csg).
There one can also find introductory material about Perceptual Control
Theory, links to other relevant sites, and instructions for joining the
mailing list on which technical discussions of Perceptual Control Theory
are conducted.
About this issue
This special issue of the International Journal of Human-Computer
Studies was solicited by the Journal Editor, Dr. B. R. Gaines, with the
goal of illustrating the breadth of application of Perceptual Control
Theory. This goal clearly extends the scope of the articles beyond the
usual realm of "Computer-Human Interaction," though as it happens, most
of the papers either use computer simulation or are concerned with
easing the human use of technology. Papers for this issue were solicited
by the Special Issue Editor only through messages posted to the mailing
list of the Control Systems Group, on which questions relating to
Perceptual Control Theory are discussed. Certain of the papers were
solicited explicitly, because they had been the subject of discussion on
the mailing list despite not having been published elsewhere. Others
were submitted in the normal way, while yet others were written
especially for the special issue at the request of the Editor.
The issue starts with a paper by Bourbon and Powers in which the three
main classes of behavioral theory-stimulus-response, preplanned action,
and control-are stripped to the bare essentials, and compared in their
ability to predict the results of a simple tracking task under
successively less restrictive conditions. A simple control model
provides almost exact prediction under all tested conditions, whereas
models based on the other foundations are successful only under
restricted conditions. Bourbon and Powers argue that the restrictions
are generic, and can properly be applied to the more complex
circumstances of everyday life.
In the following paper, Powers demonstrates Perceptual Control Theory in
application, by simulating an arm with shoulder and elbow joints,
showing that the so-called degrees of freedom problem is not a real
issue. A "Little Man" observes a moving target in a 3-D space, using
stereo vision, and attempts to keep a fingertip on the target as it
moves.
The next four papers deal with the use of Perceptual Control Theory or
the related Layered Protocol Theory in considering human interaction
with machines. Marken introduces a form of task analysis called
Percolate, built around Perceptual Control Theory, and discusses it in
the context of a satellite control system. Farrell et al. discuss the
relationship between Perceptual Control Theory and Layered Protocol
Theory in the context of an interface design and analysis tool called
LPTool, showing the use of LPTool in designing improvements in the
interface to a navigation and radio communication unit in a helicopter.
In a companion paper, Taylor et al. examine a critical portion of the
Layered Protocol Theory, called the General Protocol Grammar, which is
central to the design of interface dialogues, as well as being used to
analyze interactions between humans. In the last paper of the group,
Haakma uses Layered Protocol Theory to develop and test an improved
interface to a digital recording device for the consumer market.
Extending the scope of the discussion of Perceptual Control Theory,
Robertson and Goldstein report experiments on controlling the perception
of self-image. Self-image is a very high-level perception within the
hierarchy of levels postulated by Powers, which therefore is sustained
by many different kinds of actions in a variety of realms. Nevertheless,
Robertson and Goldstein are able to use Perceptual Control Theory to
study it in a rather clear manner.
The last paper in this issue, by Tucker et al., deals with the actions
of people in crowds. Using a simulation program devised by Powers, they
show that certain patterns of arcs and rings that often form in natural
crowds also occur in much the same way in crowds of simulated people
defined by a very small number of simple one-level control systems
relating to the perception of distance to a target person, and to the
avoidance of collisions with other people. The simulation program itself
is available as freeware for further experimentation. Beyond
demonstrating the arcs and rings, the program shows that a simulated
individual using very simple perceptual control can escape from traps by
moving away from the target toward which it is trying to go.
It is impossible in one journal issue to give a full impression of the
power and versatility of Perceptual Control Theory, nor of its precision
in those circumstances in which quantitative prediction is possible. The
intention of the Special Issue Editor has been instead to provide hints,
and to suggest that Perceptual Control Theory is a robust foundation for
both theoretical psychology and human factors engineering.
References
Held
Kohler
Powers (1973)
J.G.Taylor
