On modeling behavior using control theory
Unedited posts from archives of CSG-L (see INTROCSG.NET):
Date: Sun Dec 19, 1993 10:54 am PST
Subject: On modeling behavior using control theory
[From Bill Powers (931219.0930 MST)]
On modeling behavior using control theory
In discussions of control theory with engineering control theorists on thenet, it has become clear that engineers have developed certain ways of viewingthe problem of modeling behavior, and have settled on particular ways ofrepresenting it. To a great extent, the conception of what has to be explainedis dictated by the form in which control diagrams are customarily drawn, and bythe mathematical methods already available for analyzing systems with a certainorganization. The question of how to apply control theory to behavior thusbecomes the question not of how behavior is organized, but how the existingmethods of analysis can be used without change, together with their underlyingassumptions about what behavior is. The engineers have been arguing, in effect,that the methods they learned in school are completely adequate for theanalysis of behavior, and that PCT introduces nothing new. As long as thatbelief persists we will get nowhere. I think we should focus on the realissues, or give up.
------------------------------------------
The most common engineering assumption is that behavior is the externallyvisible effect that an organism has on the physical world around it. Theengineering assumption is that there is a "plant" in the world outside theorganism, and the objective of organismic control is to bring this plant to acertain static or dynamic state and maintain it there in the presence ofdisturbances.
The PCT view, in contrast, is that organisms know nothing directly of theproperties of the environment. All information they have about the environmentis found in the primary sensory signals resulting from the impingement ofphysical stimuli directly on sensory nerve endings, and in perceptual signalsthat represent further processing of the primary signals. As a result, all thatthe organism can control is a representation of the world in the form of neural(or chemical) signals. There is no way for the organism to determine the actualeffects of its outputs on the external world, so it has no way of altering itsown actions or organization to create systematic effects that are notrepresented as perceptions.
This immediately rules out "open loop control" as a possible model ofbehavior. If the organism can't ascertain the effects its output are having,there is no way for it to adjust its outputs to produce any particular effecton the world. All outputs have particular effects on the world, but theorganism can control only those effects that appear in its sensoryworld.
All control is closed-loopcontrol, in a model of organisms. If sensory information is lost, someinformation dependent on the output signals must be substituted if any form ofcontrol is to continue. And when such a substitution takes place, thecorrespondence between the perceptual signals being controlled and variables inthe external world is lost. How much difference that makes depends on therapidity with which the substituted information departs from the informationthat would have been received if the sensors were operating normally. --------------------------In the engineering approaches described on the net, there seems to be atendency to think of control behavior as one single process, often of greatcomplexity. The PCT approach is to consider behavior as the aggregate of agreat many simple control processes, working in parallel and alsohierarchically organized. The engineering approach is a matter of preferenceand custom, undisciplined by the actual organization of the neuromuscularsystem. The PCT approach is forced on us by the facts we know about theneuromotor systems.
We know of a great many simple control systems in the human organization,which deal with simple scalar variables, their integrals, and theirderivatives. These control systems are used in all kinds of behaviors, incontrolling all sorts of variables. The 600 to 800 control systems that controlthe musculature directly are employed by higher systems in every behavior thatinvolves overt action. Their organization remains constant while the behaviorsin which they are involved vary over the whole range of human experience, fromscratching an itch to multiplying two polynomials together using pencil andpaper.
Furthermore, behavior is clearly organized in larger units which alsoretain their character over a wide variety of behaviors. We learn specificskills like walking and speaking and typing, riding bicycles and driving cars,handwriting letters and numbers, opening doors and windows, putting thingsdown, picking things up, moving things from one place to another, throwingthings, pulling things apart, putting them together, and so on and so on intothe hundreds or even thousands of elementary control processes. Theseelementary skills remain the same even though they are used as the means ofcontrolling more abstract variables of all kinds, in all sensory modalities,under all kinds of external conditions.
Representing all these control processes as a single complex expression issimply not possible, and it is probably not possible that they are physicallyrealized as a single complex process. It isn't useful to try to find a generalexpression that will cover all these varieties of behavior and all these levelsof organization. We need to understand the details, because each control systemis a general-purposedevice that can be called upon in a variety of combinations by higher systemsof very different nature. A lower control system provides the means for ahigher system to bring one variable of experience to a specified value in areliable way, without the higher system needing to know how to accomplish thatend. The higher system simply generates a signal that is an example of thedesired perception, and the lower system takes care of all the necessarydetails in bringing that perception to the requested state.
The single-systemsingle-levelapproach is simply too limited to explain all of human behavior, even inprinciple. It requires that we either understand the entire human system, orunderstand nothing. Under the hierarchical approach, we can analyze behaviorsystematically, starting with the least units of organization and progressingto the more complex ones. This is the only practical way to model the entiretyof human behavior, and it is also probably the way behavior is actuallyorganized. -------------------------------To model behavior we must first observe it carefully. I do not see the requireddegree of care being taken in the engineering approaches. Too much is taken forgranted, too much is left out. Common-senseassumptions are used where a model is really needed. Diagrams leave outessential functions and connections. Disturbances are either ignored, orassumed to have convenient characteristics that allow them to be handled in astatistical way instead of in detail. Too many loopholes are left, too much isassumed without demonstration. The approach to PCT is defensive, not open.There is too much blind reliance on mathematics, and not enough attention paidto identifying the constants, variables, and functions with observables. Theorientation is theoretical, not experimental.
These are all defects that we have labored, and still labor, to remove fromthe PCT approach. Until control engineers take them seriously, it will be hardfor them to see why their approach is inadequate to the problem.
Best to all, Bill P.