Brain of the Firm (Beer, 1972)

Brain of the Firm (Beer, 1972)

Stafford Beer

"The main discovery of cybernetics after a history of twenty-five years, and indeed what gives it the right to be called a coherent science, is that there are fundamental principles of control which apply to all large systems... This book is entirely concerned with the contribution which cybernetics, the science of control, can make to management, the profession of control."

"What the system really needs, and all it needs, is a way of measuring its own internal tendency to depart from stability, and a set of rules for experimenting with responses which will tend back to an internal equilibrium. There is no need to know in advance what might cause a disturbance; there is no need to know what has caused a disturbance. To be aware of something happening and label it disturbance, and to be able to alter internal states until the effects of the disturbance are offset, is enough."

JLJ - cybernetics possibly reaches a high tide in 1972 with this work of Stafford Beer - a predictable fall from grace happened when the promises of cybernetics were not fulfilled, more likely due to the complexity of the world itself and a general inability to model it with enough detail to permit the kind of "scientific" management explorations envisioned by Beer.

In all honesty, you can use a spreadsheet to model most of what you need to do. Beer's ideas concerning daily 'programmes' for action and possibly alarms indicating early warnings when the system is out of whack - have value. The rest sits uncomfortably next to Norbert Wiener's out-of-control mathematics and other unnecessary reductions, of what arguably cannot be reduced. I can see very large firms, or perhaps investment portfolios, using these concepts to manage risk, or perhaps the Federal Reserve modelling the economy in order to determine effective interest rates.

Beer seems to stumble when he overlooks the fact that specific rich detail might be needed from our sensors - cues which humans easily perceive but are difficult to describe - information so important that its lack might make the whole attempt at cybernetic control unproductive. A cybernetic system cannot build relationships with other managers. A final fatal flaw, in my opinion, is the clear lack of question-and-answer in cybernetic 'thinking'... what we have here is more of a decision support system, with the word 'possibly' placed in front of the 'decisions' produced. Such a system might be able to run a factory, but it cannot scheme and plot and think outside of the box.

However, comma, his ideas remain useful for game theory. The kind of explorations Beer visioned for management in a firm become conceptually viable for machines playing complex games of strategy - or war - providing that we can generate predictive models for likely moves and extrapolate over time to establish a practical and opponent-confronting stance with viable adaptive capacity. To those who plan for establishing effective adaptive capacity, the "horizon effect" is just a disturbance that is regulated, managed and strategically controlled.

As far as other books Beer has written, I think I'll have another Beer... perhaps two...

p.31 The manager is the instrument of change (otherwise what is he doing?) which is to say his job is that of control.

p.32 The main discovery of cybernetics after a history of twenty-five years, and indeed what gives it the right to be called a coherent science, is that there are fundamental principles of control which apply to all large systems... This book is entirely concerned with the contribution which cybernetics, the science of control, can make to management, the profession of control.

p.33 The issue before us which above all needs to be considered from this standpoint is the relationship between man and machine. We have seen the computer used in the role of a fast adding machine, regarded as a quicker and possibly more accurate way of 'doing the sums'. We shall need to regard it as something more than this, and to use it far more intelligently. For, moron though the computer may be, its huge capacity to store information, its fantastically rapid retrieval capability, and its vastly suprahuman capability to juggle with thousands of quantified variables simultaneously, offer man an asset which as it were entitles the machine to an equal partnership. [JLJ - yes, but could Beer have possibly considered that man, in his infinite wisdom, would use the machine to watch cat videos on youtube, play video games, and chat on phone?]

p.35 If there are principles of control, we shall have to start specifying them... The idea is that we should really sit down and think. What is control all about? [JLJ - perhaps control is useless without taking into account ideas which support the critical success factors of a scheme. When we learn how to do something, what we are really learning is a scheme - we rarely think about this, but there are actually multiple ways to go about our business. A manager exercises control because he has the vision of a scheme, and makes decisions to move tactically or positionally, towards higher level goals where a direct path is likely blocked.]

p.35 Let us define awareness behaviourally, that is, by specifying the behaviour which is to count as typical of an aware system.

p.36-37 what is to count as a response to a stimulus? ...Clearly what is to count as a stimulus is an interference which affects the system's operation in some way... What is to count as a response is some change in the system which makes sense only in terms of the stimulus offered... A stimulus is something which alters the operation of the system. The system's response is something it does that has to be interpreted in terms of the stimulus. In general, this means that the system avoids or otherwise counteracts a stimulus which disrupts its activity, and embraces or seeks to increase a stimulus which favors its activity. [JLJ - yes, but you do this as part of a grand "scheme" for going on. If you open a store, you ought to have a business plan to run the store and general-purpose practical experience which tells you how to handle most problems. You have pre-planned answers to whatever might arise. You simply execute your scheme. If another store does this better than you do, and draws all your customers away, then you change your scheme, or else fold up shop.]

p.37 We said at the outset: 'take some kind of routine activity as exemplifying the system's natural dynamics', and this is really the clue to thinking about control. Systems are and systems work; if not, they are not systems at all. Control is what facilitates the existence and the operation of systems.

p.38-39 What the system really needs, and all it needs, is a way of measuring its own internal tendency to depart from stability, and a set of rules for experimenting with responses which will tend back to an internal equilibrium. There is no need to know in advance what might cause a disturbance; there is no need to know what has caused a disturbance. To be aware of something happening and label it disturbance, and to be able to alter internal states until the effects of the disturbance are offset, is enough. [JLJ - what the system really needs is a scheme for maneuver, involving a stance against the known and the unknown, and a way of intelligently constructing a diagnostic test of its ability to adapt to realistic attempts to knock it out of this stance. The system uses the stance to support goal-seeking efforts in a complex world.]

A system that can do this, that can survive arbitrary and unforecast interference, is known to cyberneticians as an ultrastable system (following Ashby's nomenclature).

p.47 feedback... will need to be cleverly designed to produce a damping rather than an amplification of input fluctuations. Given that it can be done, however, we have that impressive result we are seeking: a self-regulating mechanism which does not rely on understanding causes of disturbance but deals reliably with their effects... What matters is that control should be exerted regardless of the cause of disturbance.

p.51 There are three fundamental components of any control system, as we have seen: an input setup, an output setup, and the network that connects the two

p.65-66 Proliferating variety must be stopped; the potential must be cut down, and down, and down - even though we cannot prove the best way to do it... The question is: how does a system conveniently and effectively undertake this fearful task? The answer is: by organization.

p.68-69 An algorithm is a technique, or a mechanism, which prescribes how to reach a fully specified goal... An heuristic specifies a method of behaving which will tend towards a goal which cannot be precisely specified because we know what it is but not where it is... Heuristics prescribe general rules for reaching general goals... To think in terms of heuristics rather than algorithms is at once a way of coping with proliferating variety. Instead of trying to organize it in full detail, you organize it only somewhat; you then ride the dynamics of the system in the direction you want to go.

p.70 The failure to understand the role of heuristics in complex systems at large slops over into much computer thinking.

p.70-71 The subtle point is this: if the goal is not recognized in detail, an heuristic is required, so the computer must be supplied with an algorithm determining an heuristic. That is the basic trick... the computer must find out over a period, by trial and error, the courses of action which lead to better results of control... Alter the solution you are now using a little bit, says the algorithm, and compare the outcome with the erstwhile outcome. If this is more profitable, or cheaper, or whatever else we say, adopt it. Go on like this until any variation you make leads to a worse result than you already have. Then hang on to this solution, until the situation changes; whereupon you may do better once again by producing a new variation.

p.72-75 There are several important points to be made about heuristic control...

An heuristic will take us to a goal we can specify but do not know... The algorithm... specifying this heuristic stipulates the eventual discovery of a strategy... This strategy cannot be worked out in advance.
If we give a computer the algorithm which operates the heuristic, and wait for it to evolve a strategy, we may find that the computer has invented a strategy beyond our own ability to understand. This is quite possible in so far as it can make trials more quickly, more systematically, and more accurately than we can, without pausing for play or rest, and without forgetting the results...
This being the case, it is time to start recognizing the sense in which man has invented a machine 'more intelligent' than he is himself. The thought is annoying...
The argument that 'computers can do only what they are told' is correct but highly misleading. It suggests that they must remain the moronic slaves of their inventors. In fact we are telling them to learn, and giving them a training algorithm; but they learn more efficiently then we do and must pass us in the ability to achieve heuristic control.
The argument that the output of a computer is only as good as its input... is true for algorithms specifying algorithms, but not for algorithms specifying heuristics...
The mechanism we are using is precisely the old servo-mechanism discussed much earlier, in which error-correcting feedback is derived by a comparator from actual outcomes contrasted with ideal outcomes. But the outcome is measured... in terms of the whole system's capacity to improve on its results as measured in another language...
Secondly, the servomechanism's feedback does not operate on the forward transfer function as such. It operates on the organization of the black box which houses the transfer function. That is, it experiments with the connectivity... As effective structure emerges, this is what cuts down the capacity to proliferate variety.
...the conclusion (Chapter 2) that feedback dominates the outcome still holds...
[higher order control systems are possible]
[still higher order control systems are possible]
This argument continues until the hierarchy of systems... reach some sort of ultimate criterion... It can only be survival...
...the training process for here and now is the evolutionary process for the epochs ahead.
So when we said than an heuristic organizes a system to learn by trying out a new variation in its operational control strategy, we might equally have said that an heuristic organizes a family of systems to evolve, by trying out a new mutation in its genetic control strategy. The aim of adaptation is identical.

p.81 This state of affairs illuminates the need for a constant flirtation with (what we usually call) error in any learning, adapting, evolutionary system.

p.82 Error, controlled to a reasonable level, is not the absolute enemy we have been taught to think it. On the contrary, it is a precondition of survival... this point... is not understood by management... Thus it follows that when change is really understood (for some reasons) to be necessary, people resist the need, because to attempt change is automatically to increase the error rate for a time, while the mutations are under test.

p.110-111 Some people think of a model as a mathematical equation, others think of it as a theory, still others as an hypothesis, and yet others as a physical thing... In these expressions four key notions are embedded. [JLJ - formatting for readability added]

There is scaling-down in both size and complexity...
There is transfer across, whereby actual parts of actual things are represented again in their relative positions.
And arising from this there is workability, by which I mean that the model can, in principle anyway, operate like the original...
The model is a good model if it is appropriate...

In general we use models in order to learn something about the thing modelled

p.116 The nervous system is not easy to understand, and few people seem to make much effort to understand it. If we wish to make a model of it which will illuminate the problems of management structure, however, the attempt must be made.

p.122-123 A truly vital kind of filtering process in the brain stem is the 'arousal' mechanism... The risk we run by providing ourselves with all this protection is serious. If something happens which is actually dangerous, or indeed of special interest in some other way, we cannot afford to have the sense-data which recognize the relevant details filtered out like so much noise... On the other hand, if the arousal mechanism becomes oversensitive, as it sometimes does when we are debilitated, we find ourselves jumping out of our skins at any sudden noise... Now this whole business of arousal is apparently tied up with the whole business of filtering. We should not think of a filter simply as an inhibitor, whereby lots of data is stopped or reduced. It can also be a facilitator - if it allows only certain kinds of message to pass through, and stops or inhibits the others.

p.128-129 What matters to the firm's top management is not so much 'the facts' as 'the facts as presented', and the presentation chosen can govern the outcome of even the most important and well-considered decision.

p.152-153 whether we speak of respiratory control in the body or production control in a works, we are operating on the same cybernetic principles which apply to all adaptive systems... My respiration continues autonomically; but at any moment I may decide to hold my breath or to take a deep breath. If I do so, the autonomic system must subsequently cope with the consequences when my conscious intervention has ceased and I am thinking about something else. The same is true in management

p.159 It does not follow that the firm must have a whole machine of its own, because it may hire terminals fairly cheaply that are connected to a time-sharing service. Indeed, it is predictable that it will soon be a status symbol not to own a computer. [JLJ - yes, this was the thinking circa 1972. Most businesses do not own the buildings they occupy - why should a business own the one computer it uses when it can share?]

p.160 So the question: 'What is the model a model of?' is answered by: 'a viable organism' - regardless of its size.

p.161 From our point of view, at any rate, we can do no more than acknowledge that the model ought to be applied to the major entity with which we have to deal.

p.163-164 We ought at one to take note of a very special fact. Thanks to predictive techniques of modern statistical theory, it is possible for the controller (1B) to anticipate deflections from the plan and to begin modifying its instructions in advance.

p.165 There must be an initial plan. There must be a constant updating of that plan on the central command axis - otherwise the plan is no longer calculated to meet the needs of the organism as a whole. There must be immediate recognition of the actual state of affairs; otherwise time-lags are introduced... There must, finally, be a way of commanding the subsidiary to update its plan to meet whatever difficulties are encountered.

p.173 A switch is a device, or a whole mechanism, which transfers signals from one part of the system to another. We have already met many switches, in both the physiological and managerial aspects of this book. They are not simple flick-over devices, such as those which operate the electric light, in either of these contexts. We examined their nature rather carefully in Part I, and called them algedonodes.

p.179 The whole point about cybernetics as a science is that it should abstract the laws behind the control system it studies, and make them generally available. When I call the algedonode a cybernetic paradigm, I mean that it is a theoretic mechanism which accounts for the body's switchgear, and that it can be modelled in a management information and decision system under automation. Now the messages passing through an algedonode are either to be 'toned up' (perhaps to the point of acceleration and the whizz-bang declaration of their indomitable existence) or 'toned down' (perhaps to the point of extinction). In the paradigm, this means raising or lowering the conditional probability that the switch actually transmits the message.

p.180-181 The organism is carrying on quite cheerfully on its own; the higher authorities are doing something different altogether; the enterprise as a whole is asleep... It is, I think, quite helpful to regard this as a natural state of affairs. This means to say that all is well with the organism except that it is asleep - therefore the problem is to awaken it... The positive solution which the body has found is a special and definite arousal mechanism which alerts the higher brain centres to quash the upward inattentiveness natural to System Three at the lower brain levels.

p.191 in arousal... the rate of change involved is so steep... if that entails reprogramming, there is clearly no time available to embark on experimental programme modifications. There needs to be a separate programme all ready, which can be selected in short order and used.

p.198 Now this book can really begin.

p.206 The dynamics of this whole structure depend on the quantification of its performance... From the corporate standpoint, divisional performance is about both short- and long-term viability.

p.214 People normally classify activities according to their manifest appearance to the world at large... What matters to management about two entirely different products is not whether they look alike but whether they are profit-earners or not.

p.217-218 To understand what happens next (at Step 5) it is vital to realize the basic procedures for controlling activities are already settled... Step 5 is a dynamic process of adjustment, which selects particular plans for implementation, and quantifies the required programmes in a feasible form for the present epoch.

Planning therefore consists of the arrangement, within these known procedures, of a number of building blocks which are forecast actualities... in the short run, in responding at Step 5 with motor output to a sensory input, the governing required must needs be based on an accurate assessment of what will actually happen... There is a general model of capability in store, which was not too difficult to construct nor too expensive to record - because it is idealized.

p.219 Programmes for action within the division, continuously generated in this fashion at Step 5, will be assembled for issue to the operating centres as required at the synaptic Step 6.

p.245 There is just one clue in management history as to how the whole big-switching capability can be made to work in practice... The answer was the 'war room' - a large operations centre equipped with relief maps spread out on tables, on which incoming information could be depicted by the movement of counters... Some attempts have been made to model this operation in modern businesses, using electronic displays of data in place of the maps... The contemporary activity which most resembles the original war-room control centre is surely Mission Control in the Space Centre at Houston, Texas, where the real-time command of space operations is conducted.

p.248-249 But the very first question that a sensible management team would ask itself in these circumstances is: what are our alternatives? ...Today, thanks to the techniques of fast-time simulation, all manner of experimental situations can be set up in the control centre and worked through by the management team. There is no claim in this that our methods can predict outcomes exactly. Science is not a matter of crystal-ball gazing. The objective, on the other hand, is valid and valuable. It is to explore the responses of the system to various alternative actions, in order to see which areas of a problem are more sensitive than others to the assumptions which management is making. And this is done to test which policies are more vulnerable than others to a range of likely events. [JLJ - yes, yes, provided you have modelled the world properly and provided that you have selected an appropriate level of paranoia about the unexpected. The alternative here is to find a really bright individual and give him/her stock options in the company. S/he will scheme and fret and sweat to produce results because s/he will benefit directly from them.]

p.249 A keyboard interface between a senior manager and his computation facilities is sheer nonsense - because he is not a typist. [JLJ - well, maybe he should learn to type - and that idea is not sheer nonsense.]

p.290 The major thread unravelled by management cybernetics is the thread of variety - its generation and proliferation, its reduction and amplification, its filtering and control. For variety is the very stuff and substance of modern management in a newly complicated milieu... When we make our models and classify our insights in terms of variety, we perceive what management is really about - whatever the variety sources may be. At all times the management process seeks to procure requisite variety in stabilizing the enterprise towards survival. This it does either by devising methods for reducing the variety with which it is presented, or by seeking to increase its own variety, or (more usually) by doing both at once.

p.291 the best orientation of managerial structures is towards deciding what at any moment is the right form of variety reduction for the organism as a whole. So despite all our variety engineering... the most massive of the variety inhibitors is essentially self-organizing.

p.291 Readers are by now alert to the existence of redundancy as a powerful mechanism in circumstances where the organization is computing with unreliable components

p.300 The viable system is a system that survives. It coheres; it is integral. It is homeostatically balanced both internally and externally, but it has none the less mechanisms and opportunities to grow and learn, to evolve and adapt - to become more and more potent in its environment.

In all of this the viable system may succeed sensationally, spectacularly fail or it may muddle along.