INTRODUCTION TO CONTROL THEORY

William T. Powers

November 1990

Control theory, as we use the term, means engineering control theoryadapted for use as a model of the behavior of living systems. Those who alreadyunderstand engineering control theory therefore already know part of the story.The rest of the story lies in the way we organize a model of control to explainorganismic behavior. Sensors, comparators, and effectors appear in this modeljust as in ordinary models of nonliving control systems. Where we understandenough of real behavior, the models are set up much like models that others useand for the same purpose: to analyze behavior through simulations. But thereare some critical differences.

In a living control system, the reference input is not accessible fromoutside the system. Engineering diagrams commonly show the reference signal asan input from the outside world, which it is in artificial systems: it's themeans by which the human user tells the control system the level at which tokeep its controlled variable. In a living control system, the "user" is thewhole organism. Reference signals are set by higher systems that are alsocontrol systems (the higher systems act by adjusting reference signals forlower systems). In some cases the reference signals are derived from genetically-specifiedinformation (for example, the reference signal for body temperature). In themajority of the control systems that exist in the brain, however, theorganization is learned within a general matrix of preorganization, andreference signals derive from the operation of a multi-leveled,"massively parallel" system.

One of the basic insights behind our uses of control theory is that allcontrol systems control their own inputs, not their outputs. In engineering,this fact is obscured because the inputs are arranged so as to represent anexternal variable of interest to the user of the system, generally a variabledirectly affected by the actions of the system: position, temperature,acceleration, pressure, and so on. But a little thought will show that suchvariables can be known to the system only as signals generated by sensors; inevery case it is the signal, not the external variable, that is under control(just picture what happens when the sensor drifts out of calibration). Ourmodel must be understood from the viewpoint of the system itself, not that ofan external user.

The human system knows the external world through millions of sensors. Itaffects the external world, and thus its inner world of sensory signals, by itsactions. The sensory signals also play a part in the production of action: wepropose, specifically, that it is the same role played by the sensory signalsin control systems. This leads to a new understanding of behavior, in whichaction and perception are part of a closed control loop, the action serving tomaintain the perception at whatever level is currently specified by an innerreference signal. External disturbances tending to alter the signals, theperceptions, result in actions that oppose those effects, thus leading to thespurious appearance that the

system senses the disturbances and simple reacts to them.

This picture is very different from a stimulus-responsemodel, and it is also very different from a cognitive or command-drivenmodel. One level in the model does not tell a lower level what act to perform:it provides an example (in the form of a signal) of the state to which thelower system is to bring its own sensory signal. The lower system itselfprovides the action needed to match perception to the reference. A sensorysignal entering a control system does not cause any particular action to occur;the action is based not on the perception but on the DIFFERENCE between theperception's state and state currently being specified by the referencesignal.

This model is very tightly interconnected. A perceptual signal in a givencontrol system is derived from the perceptual signals in a set of lower-level systems. The derived signal is of a new type; it is a function of the setof lower perceptual signals. This higher-levelperception is compared with a reference signal, and the difference is convertedto a set of output signals. These output signals enter THE SAME SYSTEMS FROMWHICH THE LOWER-LEVEL PERCEPTIONS CAME, serving as reference signals that specify the states ofthe lower-levelperceptions. All loops are closed: all behavior at all levels is purposive.Every effect generated by any system is controlled in terms of the perceptionthat represents it: nothing organized ever happens open-loop.

The evidence in support of this model ranges from excellent at the lowestlevels to sketchy at the highest. Where we know how to do experiments, weconstruct quantitative working models and match them to behavior by adjustingtheir parameters. We're trying to expand the scope of these experiments tohigher levels, but the going is slow. One factor that encourages us is that allcontrol loops, in this model, can be detected and tested from outside thesystem, because all loops are closed, ultimately, through the environment.Where the model is wrong we can find out that it is wrong.

The model is also approached in another way, as an organizing principle forreinterpreting phenomena of behavior. Given the basic organization of controlas we see it at the lower levels, the question is whether higher levels oforganized behavior also make sense in these same terms --more sense than when interpreted in conventional ways. So far the answer seemsto be a unanimous yes. We are trying, however, to extend the method of modelingso it can be useful in areas where quantitative experiments are difficult. Inthis way we hope to test and buttress the insights of our clinician-membersand real-lifeinvestigators by linking their work to that of our computer modelers. Bothcontingents will learn from this interaction. But all have a long way to go.There are more than enough research problems awaiting us at all levels ofanalysis.

While our uses of control theory have many roots in the past and manyresemblances to the work of others, our approach is basically not connected toany mainstream line of development. It is a new departure, almost areconstruction of behavioral theory from scratch. Some of us are convinced thatit amounts to a revolution in the life sciences.

Bill Powers