Course Handout - Basics of Cybernetics
Copyright Notice: This material was written and published in Wales by Derek J. Smith (Chartered Engineer). It forms part of a multifile e-learning resource, and subject only to acknowledging Derek J. Smith's rights under international copyright law to be identified as author may be freely downloaded and printed off in single complete copies solely for the purposes of private study and/or review. Commercial exploitation rights are reserved. The remote hyperlinks have been selected for the academic appropriacy of their contents; they were free of offensive and litigious content when selected, and will be periodically checked to have remained so. Copyright © 2010, High Tower Consultants Limited.
First published online 14:18 GMT 28th March 2003, Copyright Derek J. Smith (Chartered Engineer). This version [HT.1 - transfer of copyright] dated 18:00 14th January 2010
Earlier versions of this material appeared in Smith (1997; Chapter 4). It is simplified here and supported with hyperlinks.
1 - Control Engineering (pre-1948)
One of Humankind's greatest problems is the relative weakness of the human body. Ever since civilisation began, humans have struggled to supplement their physical strength by external means wherever possible. They have used draught animals to pull their ploughs and carts, and variations on the "mighty five" simple inventions (the lever, the wheel, the inclined plane, the screw, and the wedge) to move their obstacles and erect their buildings. But the bigger the weight involved, the greater the risk of accident and injury. All power sources - be they your own muscles, harnessed oxen, rushing water, wind, steam, or whatever - require control, and the more powerful the power source, the finer that control needs to be. The name given to the resulting technology is control engineering.
In the early days, of course, control was exercised by the person using the mechanism in question. For example, when the millers of yesteryear needed to turn their mills back into the wind every time it shifted, it involved manhandling several tons of wood and iron. And they had to put up with this for about 600 years before Edward Lee, in 1745, developed the mill "fan-tail" to do the job automatically for them [picture]. This mechanism consisted of a small rotor mounted at the rear of the mill cap, and set at a horizontal right angle to the main sails. As long as the wind struck the main sails directly, the fan-tail stood idle. When the wind veered to one side or the other, however, the fan-tail started to rotate one way or the other, and drove gears which cranked the mill cap back into the wind again.
There was a similar problem controlling the speed of rotation of the millstones. If this got too high, then the driven stone tended to bounce around against the fixed stone and the quality of the grain suffered as a result. Again, adjustments could be made manually - a process the millers called tentering - but on a gusty day this would be a full-time job in itself. This problem was eventually solved by an automatic control system introduced by Thomas Mead in 1787, and relying upon the use of a "centrifugal governor" [picture]. This was a double conical pendulum, that is to say, a pair of fly balls suspended on hinged arms from a spinning base plate (a contrivance known about for some time previously - but in miniature - to clockmakers). The base plate was attached to the millstone drive shaft (so that its speed varied with the millstone speed), and the faster it turned, the further out the fly balls would fly under the influence of centrifugal force. It remained only to connect the outward swing of the fly balls to some means of slowing the main drive and you had a means of making continuous and automatic adjustments to your speed. This adjustment was achieved either (a) by physically pressing against the turning millstone, or (b) by variably "reefing" or "feathering" the slats which made up the sails (or on occasions, indeed, by a combination of the two).
As it happened, the textile millers were having similar problems at about the same time. Theirs came from the fact that the act of spinning requires the tension on the threads to be very precisely maintained. Yet the steam engines which were being installed to drive the spinning machines were very difficult to control. When you put more coal in the furnace they went faster (about twenty minutes later), and when you put them to work they went slower (because you used up all the steam). And when they went faster they snapped all the threads, and when they went slower the quality of the thread became lumpy. The textile millers had been getting by for some time by employing men to adjust the steam pressure manually, but this was expensive, and their complaints to the builders of the engines prompted one of them - the great James Watt himself - to note how effectively Mead's centrifugal governor had coped with the flour millers' problem. It took Watt only a year to install a similar device on a steam engine (1788), and it soon became the control mechanism of choice. As the machine came under load, the speed decreased, the fly balls dropped, and more steam was automatically supplied. If the machine got too fast, on the other hand, the steam supply was just as automatically eased off.
A good formal definition of this sort of governor is given by the physicist, James Clerk Maxwell:
"A governor is a part of a machine by means of which the velocity of the machine is kept nearly uniform, notwithstanding variations in the driving-power or the resistance." (Maxwell, 1867:270.)
2 - Control Loops and Feedback
In the preceding examples, the overriding principle is that of control, and to have control you have to be able to manipulate information. Specifically, you need to transmit information on how you want your mechanism to perform in the first place, then you need to receive information on how things are currently going (in order to keep that performance within tightly preset limits of acceptability), and then you need to transmit information whenever you want to make the necessary adjustments. This gives you two types of control information. The transmitted information is feedforward - the information which instructs a mechanism or process on what needs to be done, and the received information is feedback - information coming back from that mechanism or process, and telling you how things are progressing.
These two fundamental types of information then circulate in a very special way in what is known as a control loop. This is a mechanism (such as Lee's or Mead's) for detecting deviations from some preset standard, and for taking appropriate corrective action. The elements of a typical control loop are as follows:
(a) Standard Setting: This is where a planned action is initiated in a given device by giving it a flow of initial command information (the feedforward described above). This sets the level of required performance. In a thermostat, for example, it is where the dial on the wall is turned to the desired temperature, in a steam engine it is where the centrifugal governor is adjusted to deliver a given output r.p.m., and in a single-handed yacht it is where the automatic steering gear is set to maintain a certain angle to the wind while the yachtsperson goes below to get some sleep. The unit which stores this standard is called the "comparator".
(b) Output Monitoring: This is where the ongoing progress is somehow measured, to make sure that things are going to plan. The unit which carries this process out is called the "sensor", and the information it senses is "fed back" to the comparator for evaluation.
(c) Comparison: This is where actual performance is compared with required performance. It takes place in the comparator, and generates, in turn, what is called an error signal. The error signal shows the extent of the divergence of actual away from required performance, and is the basis of the corrective action which then needs to be taken. In our thermostat, for example, the error signal might be saying that the room you wanted kept at 65o was actually at 62o.
(d) Compensation: This is where remedial action is taken by the device controller to reduce the error signal. This remedial action can take many forms. It might simply be to switch a process on or off in its entirety (which is what your living room thermostat does to your central heating boiler). Or it might involve something slightly more complicated, such as fractionally turning a device up or down. Or it might involve something even more complicated still, such as arranging for the resulting corrections to be roughly proportional to the size of the error signal. It is also possible to reduce the error signal by asking for the standard to be adjusted, but this requires extremely complicated logic.
(e) "Hunting": This is an inability on the part of a control system to establish a smooth level of performance. It is where output "pulsates" above and below the required performance level, typically as a result of "shortcomings in the governor or other speed control systems" (Chambers Science and Technology Dictionary). It occurs because there is always a time-lag between the detection of an error and the application of the correction. Moreover, if the correction is not carefully applied it is easy to over-compensate. Thus, if you need to steer left but overdo it, you will have to correct to the right; whereupon you might overdo it again and have to go left a second time; and so on, and so on, until either you learn to steer more gently or you end up hitting something.
Figure 1 shows these elements at work in a simple control system. Look for feedforward, feedback, and the various elements of the control loop, and note especially that the whole process is self-regulating - you switch everything on, set the standard, and then leave it to look after itself.
Figure 1 - A Simple Control System: The standard, or "required", level of performance is stored in the comparator (shaded, lower centre). Performance is then regularly sampled by the sensor (shaded, right), and compared with the stored standard. If a discrepancy is detected, an error signal is passed to the device controller (shaded, upper centre), and suitable corrective action taken. Note the generally clockwise direction of the feedforward and feedback aspects of the cycle. Note also how important it is to distinguish the occasional external manual elements of the process (beyond the left system boundary) from the continuous internal automatic ones (between the system boundaries). A considerably more sophisticated version of this system is shown in Figure 2.
Enhanced from a black-and-white original in Smith (1997; Figure 4.2). This version Copyright © 2003, Derek J. Smith.
3 - Negative vs Positive Feedback
In fact, two types of feedback need to be identified, namely negative and positive. Negative feedback is where corrective action is taken to reduce, or "damp", the amount of an error. This is the sort of feedback which gives us the classic "closed loop" control system described above. It also gives us the feedback we are already familiar with in biology under the name homeostasis (Cannon, 1927). Positive Feedback, by contrast, is where the correction is made in the same direction as that of the original displacement. Each pass around the feedback cycle thus magnifies the displacement instead of diminishing it. This means that we can no longer refer to the displacement as an "error", because not only do we want it to be there for some reason, but we also want it to be bigger than it already is. Positive feedback is only needed in control systems if the amplification of a signal is needed. Indeed, neurophysiology itself only started to make real advances with the coming of positive feedback electrical amplification systems capable of sufficiently magnifying the very small neural currents involved. Without positive feedback we would never have had the EEG nor the microelectrode.
4 - The Servomechanism
Another important control concept emerged with the development of power-assisted steering for ships. Here, too, the essence of the problem was that muscle power alone was not enough. The earliest sailing ships, for example, were steered by trailing an oar over the stern. It then occurred to people to fix this oar in place on "hinges" so that it became a rudder (Pliny attributes this invention to Tiphys, helmsman of the Argo, and a member, therefore, of Jason's famous argonauts), and this arrangement was then improved by linking the rudder by pulleys and ropes to a steering wheel placed more conveniently up on the bridge of the vessel. The turning of all such rudders required considerable manual effort especially in high seas, and in vessels of any size block-and-tackle systems had to be used to "gear down" a lot of turns on the helm (ie. the steering wheel) to a small displacement of the rudder itself.
With the coming of steam, came the opportunity for steering systems to be power assisted. This allowed greater speed of response to the helm, and thus greater manoeuverability and safety. The first steam-powered steering system was patented by Gray in 1866, and installed in the SS Great Eastern. The essence of the invention was that when the helm was moved to a new position no direct attempt was made to move the rudder. Instead, the helm moved a steam valve which powered a gear-train which moved the rudder. A manual pressure on the helm of just a few kilograms applied a fully controlled and highly demand-sensitive pressure on the rudder bar of many tons. And the clever part came in shutting that power off automatically when the rudder reached the desired position (at which time the physical displacement of the rudder had "caught up" with the steersman's intent).
In 1868, the French engineer Jean Joseph Farcot patented improvements to this sort of control machinery, and in 1873 published a book entitled Le Servo-Moteur, thus introducing the word servomechanism into the control engineer's vocabulary. Here is a formal definition:
"The 'servo' notion implies that heavy work is done by a strong slave under the direction of a master, who by virtue of this assistance need exert no effort beyond that of formulating commands." (Roberts, 1978:116.)
Servomechanisms (or "servos", or "slave systems") are important because they allow a small control system to control pieces of far heavier machinery. Thus the helmsmen of the late twentieth century can easily have one-finger control over the steering of a supertanker - an amplification factor of about 1:1 billion. And much the same sort of thing happens in biology, because our brains are small control systems and our bodies are pieces of far heavier machinery. In fact, neurons weighing in the order of 0.00000001 gm can control a body weighing around 100,000 gm - a amplification factor of 1:1,000 billion. (Or even more, should the neuron happen to be controlling the finger which is steering the supertanker!) This being so, it helps if we are able to recognise servomechanisms "hidden" away inside more complex systems, because it will give insight into the modular organisation of those systems. There are seven distinctive features to look for when doing this:
* A servomechanism's control standard is set from some distance away.
* A servomechanism has local access to the power it needs to do its job.
* A servomechanism can hold indefinitely to a given setting, but .....
* ..... ideally, it should be able to "interrupt" its higher controller should it start to experience significant difficulties in holding that setting.
* Ideally, it is sensitive, quick to respond, and properly damped, so that it responds smoothly and without hunting (see earlier this chapter).
* Ideally, when told to seek a new setting it applies its changes proportionately to the size of the change required, that is to say, big changes are implemented at a greater rate of change than small changes.
* Ideally, it is "user friendly", that is to say, it is richly equipped with built-in safeguards.
It is also useful to study the distribution and use of memory resources in a control system. We have already mentioned that a comparator cannot actually do its job at all without some sort of memory capability, and things get worse in servo-assisted systems due to the physical separation of servo and controller. This point can be particularly clearly seen in systems which make use of efference copy and reafference - terms introduced by Von Holst and Mittelstaedt (1950) to describe a clever way of coping with the potentially excessive feedback found in most biological systems. The problem is that motor activity - efference - does not just induce movement, but also affects what is picked up by the senses. As soon as you start moving, proprioceptors will tell you about limb position and balance, cutaneous receptors will tell you about changes in touch, pain, temperature, and pressure, homeostatic systems will signal requests for blood pressure and blood glucose maintenance, and special senses (eyes and ears) will detect the changing visual and auditory shape of the world. The senses, in short, are involved every bit as much in motor activity as are the motor pathways.
What efference copy systems do, therefore, is subtract what you expect your senses to tell you next from what they actually tell you next. This gives zero if things are going to plan, but a non-zero error signal if they are not. This comparison is achieved by momentarily storing an image of the main motor output - the efference copy - and by then monitoring what is subsequently received back from the senses - the reafference. The principal benefit, of course, comes when the two flows totally cancel each other out, because this leaves the higher controller free to get on with more important things. Every now and then, however, the system encounters some sort of external obstacle, or "perturbation". This causes the reafference not to match the efference copy, and this, in turn, causes the higher controller to be interrupted with requests for corrective action. And because the sensors in an efference copy system thereby become capable of confirming for themselves that the effectors are working to plan, this is a highly efficient way of reducing unnecessary network traffic. Here are the formal definitions:
"[Efference copy is] basically a copy of the motor output [and] in some way represents the expected pattern of sensory messages that would be produced in the absence of external interference with the action of the effectors. [Reafference is] the message returned to the nervous system, in consequence of the issue of the command." (Roberts, 1978:141.)
An even more advanced way to make use of past experience is to develop some form of predictive control system, that is to say, a system where the efference at every level of the control hierarchy is prepared well ahead of actual need (up to seconds ahead in many cases, but hours or even days ahead in the case of more "strategic" predictions). Indeed, there is considerable insight to be gained (even as non-technologists) by considering the problems faced by robotics engineers. Maravall, Mazo, Palencia, Perez, and Torres (1990), for example, are among the many research teams working in this area, and have obtained good results with robots capable of constantly making guesses at what is coming next. Figure 2 shows these advanced concepts at work in a schematised servo-assisted system.
Figure 2 - A More Sophisticated Control System: This is a more powerful version of the control system shown in Figure 1. The three original modules are still there (albeit the internal logic is no longer shown), only now they have to support multiple muscle groups (top right), which, properly coordinated, allows the organism to develop significantly more complex behaviours such as swimming, running, flying, etc. There are also several important new modules (and important new memory resources to go with them). (1) a higher order controller (far left) replaces the external manual source of command information. This means that there is no longer any high-side system boundary, making the new layout self-controlling. That is to say, it is now capable of willed behaviour, or "praxis" (for more on which see Chapter 6). (2) multiple sensor devices (bottom right) allow more complex sensory monitoring of the outside world, but only at the price of having (3) some sort of categorising process (bottom centre) to prevent the higher order controller being overloaded. At the same time, an important subset of the output devices has to be devoted to sensor orienting (4), that is to say, to keeping the sensory devices pointed in the right direction for maximum sensitivity. The output devices need also to be equipped with a contention scheduler (5) (top centre) to prevent different devices working at crossed purposes. Fortunately, an efference copy system (6) (at hub) allows traffic from right to left to be minimised. Predictive control is provided by components at all levels. The higher order controller "thinks ahead" (7), the device controllers deploy a repertoire of prelearned skills (8), and the peripheral systems know from their efference copy what is expected next (9). Note that this diagram, if redrawn to Yourdon notation and rotated 90o to the right (so that the system boundary is at the bottom), has many points of congruence with Frank's organogramm and Smith's (1993, 1996) six-module model of cognition. Note especially that Denmark Technical University's Rodney Cotterill (eg. Cotterill, 2001) is building a convincing theory of consciousness based around the brain's efference copy system - for more on this, see our e-paper on Short-Term Memory Concepts in Computing and Artificial Intelligence, Part 5 (Section 3.12).
Enhanced from a black-and-white original in Smith (1997; Figure 4.3). This version Copyright © 2003, Derek J. Smith.
5 - Cybernetics
It does not often fall to an individual scientist to be the creator of a new science, but in 1948 the engineer Norbert Wiener gave a name to the emerging science of control. The name he chose was cybernetics, an anglicisation of kybernetes, the Greek word for "steersman". (Which, incidentally, is also - via the Latin gubernator - the root of the English word governor, making it doubly appropriate for Mead's centrifugal governor to have been called what it was.) The key publication was Norbert Wiener's 1948 book "Cybernetics, or Control and Communication in the Animal and the Machine" (although Wiener actually credits the French physicist André Ampère with having discussed cybernétique principles as early as 1834). By its nature, Wiener's new science automatically embraced the concepts of feedback, control loops, servo-mechanisms, information, communication channels, and systems theory. Porter (1969) provides a decent formal definition:
"Cybernetics is concerned with the communication and manipulation of information and its use in controlling the behaviour of biological, physical, and chemical systems. It is the basic science underlying the processes of behaviour in biological systems." (Porter, 1969:vii; emphasis added.)
Key Concept - Cybernetics: In fact, cybernetics has both a strict definition and a variety of looser usages. Very strictly speaking, it is the science of guidance. Less strictly speaking, it is the science of control in general. It thus has applications in a variety of areas, such as sociology and commerce (not relevant to this course) and biology (directly relevant to this course - see Chapter 2, etc.). Generally speaking, however, it is anything remotely connected with computing or robotics, and because it sounds such a nice word it is especially popular with writers of science fiction. Thus we have the cybermen in "Dr Who", cyberspace in "Red Dwarf", the cyborgs in the "Terminator" movies, etc. ad nauseam.
And why does all this matter? Because cybernetics tells us how to control complex, hierarchical, and distributed systems, and because the biological nervous system is a complex, hierarchical, and distributed system. The following organisational rules, for example, apply to all sophisticated control systems, including biological ones:
* There will be peripheral and central modules. The peripheral ones - the sensory and motor systems - are those which interact directly with the environment, and the central ones are those which do not.
* Many of the peripheral modules will be involved in some form of servomechanism.
* Many of the central modules will be controllers (of the peripheral ones).
* The trade-off between modular complexity and network complexity will need delicate tuning.
* As many decisions as possible will need to be made within the servomechanisms, thus avoiding unnecessary traffic to and from the controller. There will therefore be especially heavy traffic between sensory and motor devices at the periphery (just as with the physiological spinal reflexes).
* There will be coordination of the controllers (by some sort of "super"-controller). (Remember that Chapter 6 is dedicated to hierarchical control systems of this sort.)
* There will be integration of the action of the servomechanisms, in order to prevent different devices working at crossed purposes. (Consider what happens if, in your car, you press the accelerator and brake pedals simultaneously!)
* Comparator processes will require memory to hold efference copy instructions pending the arrival of the reafference.
* Control processes will require memory to allow efference instructions to be packaged up into skills.
* Supercontrol processes will require memory to allow continuity of purpose at the behavioural planning level.
Cannon, W.B. (1927). The James-Lange theory of emotion: A critical examination and an alternative theory. American Journal of Psychology, 39:106-124.
Cotterill, R.M.J. (2001). Cooperation of the basal ganglia, cerebellum, sensory cerebrum, and hippocampus: Possible implications for cognition, consciousness, intelligence, and creativity. Progress in Neurobiology, 64(1):1-33.
Maravall, D., Mazo, M., Palencia, V., Perez, M.M., and Torres, C. (1990). Guidance of an autonomous vehicle by visual feedback. Cybernetics and Systems, 21:257-266.
Maxwell, J.C. (1867). On governors. Proceedings of the Royal Society, 16:270-283.
Porter, A. (1969). Cybernetics Simplified. London: English Universities Press.
Roberts, T.D.M. (1978). Neurophysiology of Postural Mechanisms (2nd Ed.). London: Butterworths.
Smith, D.J. (1997). Human Information Processing. Cardiff: UWIC. [ISBN: 1900666081]
Von Holst, E. and Mittelstaedt, H. (1950). Das Reafferenzprinzip. Naturwiss., 37:464-476.
Wiener, N. (1948). Cybernetics. Cambridge, MA: MIT Press.