Perceptual Control Theory is described, and references are available at the Control Systems Group web site. The central motto of Perceptual Control Theory is "All behavior is the control of perception." In this context, "perception" has a technical meaning: the value of some variable internal to the brain that corresponds to some function of physical (or other brain) variables.
If it is true that every action is performed to control one or more perceptions, then to help must be to ease that control.There are many reasons why help may be useful, and together with the way help is provided, these reasons can lead to a taxonomy of kinds of help. In what follows, the one needing help is called the "user", and the one providing help the "helper."
All forms of help have one common problem, which under different circumstances may be important or insignificant. It is that most people have a high-level reference to perceive themselves to be adequately competent and capable, and a separate reference to perceive others as seeing them to be competent and capable. To use a helper would be to reduce those perceptions of competence, causing error in the relevant ECSs. One way for the user to counter the disturbance would be to reject the help, even when the help might improve the user's control of the perception for which help is available. Whether this happens depends on the relative gains of the perceptual control systems concerned with the task and with the perception of competence, and on the magnitude of the errors in each that are consequent on the use or non-use of help.
If the user can indeed perform the task almost as well without as with help, then help will normally be rejected unless its acceptance can serve as part of the control of other perceptions, such as those dealing with social relations between the user and the helper. On the other hand, if it is important to the user that the task be done well andwithout help it will not be done well , then help will be accepted despite any possible errors induced in the control systems that are concerned with competence perceptions. There is an intrinisic conflict if the user has some competence at the task, but not enough to do a satisfactory job without help. This kind of problem is manifest in many ways. It can lead to the infamous "Not Invented Here" syndrome, to climbing Mt. Everest without oxygen, and to choosing to learn by trial and error rather than from a teacher.
Even though there is an intrinsic conflict in accepting help, often the user will do so. The kinds of help can be analyzed according to the reasons the user may have difficulty controlling some percept, and according to the ways in which help is proferred. The first division of the taxonomy is based on whether the user has a current problem ("on-line" help) or an ongoing problem that can be resolved by learning how to achieve the desired control ("off-line" help). Does one give a starving person a fish, or show him how to catch fish? On-line help staves off starvation now, off-line help keeps starvation away in the long term. (It is quite possible, within a control hierarchy, for help to be on-line at one level and off-line at another; by catching a fish now, the helper gives the starving person a fish (on-line help), and demonstrates how fish are caught (off-line help)).
The various ways help can be provided on line, as described above, can be tabulated in three groups: tools, slaves, and autonomous help. Machines can perform the helping functions in the first two groups, and humans can perform functions in the second two groups. Both humans and machines can act as slaves.
| Kind of Help | Benefit | Problem |
| Tools | ||
| (Figs 1,3) Output force or gain | Increased range of control | None |
| (2,3)Perceptual precision augmenter | Increased accuracy of control | None |
| Quasi-tools | ||
| (4) slave control | Allows diversion of resources to control of other perceptions | If the user's reference levels change, the helper must be notified. User must monitor the CEV |
| (5) Parallel slave control | Augments user's effective gain and/or perceptual precision | Each actor's actions disturb the other's attempts to control. Can lead to instability. |
| Autonomous help | ||
| (6) Parallel helper control | For states far from the reference value, aids both the range of control and the effective perceptual speed. | For states near the reference, the "help" acts as an external disturbance, and stiffens the CEV against the user's actions. |
| (7) Collaborative control | Allows user to control a high-level perception by stabilizing some lower-level component near a useful value. | As with parallel control, if the user has an independent reference level for a perception correlated with the CEV acted on by the helper. |
| (8) Mutuality | Helps both partners control | Partners might separate, since the helping function is inadvertent. |
Thus far, we have discussed the on-line enhancement of control by a single ECS. Even in the case of collaborative control, there is only one ECS being helped. It is the high-level one to which the perception of the CEV affected by the helper's control actions contributes. And it is helped in the actual execution of its control function at the time control is needed. None of the help functions considered so far address the kind of help that eases the user's control at a future time. This is the kind of help provided by the "Help" facility available on many computer systems. It may be called "teaching" or "guidance."
One interesting aspect of Perceptual Control Theory is in the way "knowledge" and "skill" is represented. Conventional Artificial Intelligence deals with localized representation, usually in the form of symbols. In contrast, a neural network represents knowledge in a distributed fashion, as a pattern of weights over an entire network. In the neural network, no location can be asserted to have any individual responsibility for any item of knowledge, whereas in the AI system, the value of a symbol or a formula specifies a discrete item. Change it, and that item changes by itself, though others may change as a logical consequence. Change a weight in a neural network, and quite possibly all knowledge held in the network changes microscopically.
A Perceptual Control hierarchy combines both forms of knowledge representation. The value of a perceptual signal represents the current state of the perceived world in respect of one specific item, akin to the value of a formula in an AI representation. The formula is the perceptual input function (PIF) of an ECS, and it can be of any complexity whatever, as in an AI system. But the behaviour of the hierarchy depends on the weights and signs of the linkages among the perceptual functions and among the output and reference functions. In this, it is akin to a neural network. Furthermore, if the PIFs are redundant across a set of ECSs, they can be seen as a system that coarse-codes the sensory input, another powerful feature of some classes of neural network.
Learning in a simple perceptual control hierarchy can involve the individual PIFs changing to match some feature of the environment, it can involve changing the patterns of weights in the connections among the levels of perception, or it can involve changing the connections of output to reference. PCT recognizes two distinctly different forms of learning, continuous parameter modification (such as gradient search within the neural network consisting of the linked PIFs), and relinkage or reorganization of the network structure itself.
Reorganization occurs when control fails. It is a random process in that there is no way to determine which links will have the desired effect when added, eliminated, or changed in sign. The likelihood of reorganization in any region of the control hierarchy increases with the degree of control failure, or, more generally, with sustained deviation from desired values in what are known as "intrinsic variables." Intrinsic variables are for the most part considered to be those aspects of life that are not sensed directly but are necessary for survival, such as the states of body chemistry; but among the intrinsic variables the global state of control error is often included.
Because reorganization tends to happen when control fails, a hierarchy that develops in a stable, predictable environment will reorganize only to the extent that it can handle that specific environment. Slight changes in the environment may lead to loss of control. In the context of a computer interface, for example, a user may learn a pattern of commands that achieves a desired result, but if a file referenced in those commands does not exist on some occasion, the technique may fail, and the user is unable to control the perception (the desired result).
Based on this idea, one apparently perverse kind of off-line help is to vary the environment in such a way as to cause the user loss of control, and thereby to induce further reorganization. By forcing the user to control in a variety of enviroments, the user is likely to become a more stable and robust controller. In other contexts, this is known as "the school of hard knocks," but what it does is to develop higher-level control structures. In other words, the user develops deeper and more useful models of the system being learned.
In the school of hard knocks, the knocks should not be so hard or so frequent as to prevent control in the changed environment from being regained by simple changes in the hierarchy, or reorganization may become disorganization, inability to control even in the original environment. Random environmental variation can lead to creative reorganization. The user may learn techniques not known to the helper who disturbs the environment to induce reorganization.
Other off-line helping techniques may be categorized by one major distinction: does the technique aim to alter the way the user perceives the environment or the actions the user takes in controlling existing perceptions? A secondary distinction is whether the technique aims at smooth changes in the user's control structures (guidance) or abrupt changes (teaching). Smooth changes are more likely in changing what is perceived, and abrupt ones (doing something quite different) in changing the actions ao as to improve control of existing perceptions. But all four corners of the two-by-two matrix can occur. To generate new PIFs in new ECSs requires that the hierarchy change abruptly in both perception and action linkages. "Do it more gently" proposes a smooth change in action.