Cybernetics

Operational Research

and Automationby F. H. George, Ph.D., M.A., F.R.S.S.

A Paper presented at the Production Conference, Olympia, on 15th May, 1958.

Dr. George, who is Lecturer in the Department of

Psychology at the University of Bristol, was educated at

Taunton School and Sidney Sussex College, Cambridge,

where he read Moral Sciences and Mathematical Triposes.

He gamed his B.A. at Cambridge in 1948 and his M.A.

in 1952 and, also in 1952, gained a Ph.D. for work on

mathematical and logical models and their applications in

learning theory.

He was Visiting Professor at Princeton University, U.S.A.,

in 1953 - 1954, and carried out research at McGill

University, Canada, in 1953.

His principal interests are now

in experimental phychology,

mathematical logic and

philosophy of science, and the

synthesis of these three

disciplines in cybernetics.

MY plan is to try to set out a clear and generalpicture of what cybernetics and operational

research are about, and then to indicate their relationto automation. As far as automation is concerned,I hope to show that there is a great need to under-stand the fundamental principles on which allautomatic control is based, since it represents a verygeneral scientific methodology with far-reachingconsequences; a failure to understand these principlesis the surest guarantee of a failure in application.However, I wish to make my appeal to practicalpeople who are primarily interested in productionon the factory floor or in other everyday circum-stances, where elaborate theories, in themselves, arenot enough and so I hope to dwell on examples thathave a practical and intuitive appeal ; this factprohibits the possibility of careful proof, althoughsuch a proof of most of my statements, I will claim,is always available; and where we are explicitly con-jecturing I shall note this as a matter that requireseither proof or practical demonstration.

Cybernetics is a general name for the study of allcontrol and communication systems. Therefore, ifI study computers, either analogue or digital, becauseI think I see in them the hope of an analogue tohuman behaviour, then I am indulging in cybernetics.If I build a special purpose computer as an analogyto any process whatever, I might also be said to bedealing with cybernetics; it is clear that there are noprecise limits to the subject; it cuts across all theconventional boundaries and includes engineering,mathematics, biology, psychology, and so on and soforth. So I may be within the scope of cybernetics if

634

I simply sit down and work out mathematical equa-tions and descriptions. However, there are certainsalient features of cybernetics; the first is the idea ofcontrol being automatic and being primarily a processof negative feedback; secondly, there is the idea thatsystems may have the same properties, whether theybe organic and called " living", or inorganic andwhat we normally tend to call " machines ".

Operational research is in many ways more vaguelydefined in its limits than cybernetics and certainlysomething that cuts across it, and might even, in asense, be said to include it. It involves primarily theapplication of a strict scientific procedure to allproblems, and specialised because it supplies thecharacteristic model-theory treatment of science, tosituations which do not already come within anaccepted branch of science. In brief, operationalresearch represents the application of scientificmethod to everything and anything; we must saysomething, then, about scientific method.

Automation, which we will come to last of all, isapplied cybernetics, and perhaps something thatgenerally goes hand in hand with operationalresearch.

I would like to emphasise that there is no specialreason at the moment for trying to use these threenames in a precise way, since they are commonlyused to cover whole sets of ideas, concepts, andoperations, so let me rather try and indicate theirbreadth and importance. It is in this respect that Iexpect automation — in view of its great publicity —is already well-known, and needs less explanationthan cybernetics and operational research.

Scientific methodI must now turn to a description of scientific

method; what a scientist normally employs. Naturallythere is still some dispute about these matters, and Iwill try and give an outline in terms that willemphasise those parts on which we should all agree,since for practical applications, fine differences, aboutsuch things as the metaphysical background, forexample, I shall assume make no difference at thisstage of application.

Perhaps I can best illustrate what scientists do bylisting the characteristic steps they follow, and thefindings which they derive from these steps.

It is known that science depends on observation.We have to be able to observe things, whetherdirectly or indirectly. To do this we may use micro-scopes or telescopes to improve our range of observa-tion, but we still have to have direct contact withcertain states of certain systems.

From these observations, generalisations are madewhich can be put in the form of axioms or postulatesthat can subsequently be used for derivingconsequences and so allow prediction. The step fromparticular observations to guessing at the generalisa-tion is called the inductive step. This step may betaken on a number of different grounds ranging fromsheer guesswork, where no evidence or justificationcan be adduced in support of the hypotheses orgeneralisation, to good evidence based on a numberof careful observations. We might see a red apple,

and then another red apple, and then another3 andmight reasonably say " 1 believe all apples are red ",and since we mean this to be a statement that isempirical and capable of being tested (by observation)we must go ahead and see whether or not allapples are red. If not, then our generalisation isfalse, although if almost all are red, we might statea weaker, statistical, generalisation; " nearly all applesare red ". We shall, of course, be careful here tonotice that its redness cannot be a definingcharacteristic of an apple, otherwise there is no needto test our statement; it would be true by definitionand everything that looked like an apple and was,in fact, green would have to be called a " bapple " or" gaPple " o r something else.

Having achieved a set of statements that arewell-confirmed (to which we have not observed anyexception), then we may deduce particular conse-quences. That is " If all apples are red, then if youhave an apple in your pocket, it is necessarily red ".So, roughly speaking, the scientific method is madeup of the two processes of induction and deduction,fitted together. These processes link the two opera-tions of testing, which depend on observation andinclude careful scientific experiment, with that ofgeneralising, where the generalisations are sometimescalled hypotheses (if highly confirmed, laws) thatfollow from accurate observation statements.

Probability and statisticsIn our search for models, both in specialised

sciences and now in any situation where the scientificmethod will be employed, we shall find ourselvesusing probability theory and statistical methods basedupon probability. These can be used in two differentways which we should distinguish.

The Laplacian theory is the probability theory mosteasily applied. This says that the probability favour-able to some specified event A is given by ratio ofthe number of ways which the favourable event Acan occur to the total number of ways that the eventcan occur at all. To take the simplest sort of example,we shall say that the probability of throwing an evennumber, with one throw of an ordinary die is | ,since there are three ways of the favourable eventoccurring : we may throw 2, 4 or 6, and there aresix ways in all that the event can occur : 1, 2, 3, 4, 5,or 6, and therefore the Laplacian probability gives3/6 or £.

We can now consider sets of events with whichwe can associate probabilities in a priori fashion.Alternatively, we can count the number ofoccurrences of events and make some prediction onthe basis of what we count. Thus if we wish to knowwhether height and weight are highly correlated inany group, we can take a random sample of the groupand calculate a correlation coefficient. We could thencalculate a correlation coefficient for another groupunder the same circumstances and we can compareone fnroup with the other. These methods requirecareful application, the use of control groups, propermethods of sampling and the careful consideration ofmany other factors besides. This use of statisticsallows us to carry out an extended reasoning process

635

about groups of things or people, with some degreeof probability. We can by such methods makedeductions and therefore predictions.

The second way in which probability will enterour considerations is where we use it to describesome system. The probabilities arise either becausethe system is too large to handle any other way —this will frequently occur in industrial problems —or where we do not know all the necessary facts inadvance of our computation. This is characteristic ofinduction, and also of the human being in the data-processing which we associate with the activity of thehuman brain.

It is because of these two factors that many of ourmodels — the symbolic and abstract models used inoperational research — are of a probabilistic orstatistical kind. We shall be referring from time totime to the specialised statistical methods which weshall continually meet throughout the application ofcybernetics and operational research to industrialproblems, both at the machine level and at thebroader organisation levels.

So much for our summary of scientific method as itseems to operate and, of course, ignoring all thevarious complications in designing experiments andthe full methodology, involving detailed statisticalapplications and so on and so forth.

Language and logicWe can now regard this scientific method a little

more carefully as a relation between ordinary state-ments, in English or French or any other everyday-language, and their underlying logical structure. Sincethe structural relation between precise statements isvery important, it will often involve the use of logicalor mathematical models, or even physical models, withspecific geometrical properties. All informal descrip-tions of scientific phenomena are approximate andconvenient, but sometimes too vague, and we maythen expect to have to go over to a more precise,although equivalent, description. This sort of thingcan be exemplified by the difference between aninformal description of a scientific system and amathematical description, or the difference betweena rough description of a machine and a detailed andprecise one. We shall say about this that we have amodel of our ordinary language and this may be aprecise linguistic or logical model, or it may be aphysical model, like a. wind-tunnel is a physicalmodel of the aerodynamic characteristics of certainaircraft. These models can be regarded as analoguesof systems and if we build a model first we canobserve the characteristics of the model withouthaving to build the full-scale system. This last ideastands at the very base of operational research andthe link with scientific method is immediate, since itis directly derived from the methodology of science.

The first part of our description of scientificmethod is primarily concerned with cybernetics, sincewhereas we can normally build a machine to performthe deductive operations, an inductive one has beenmore difficult to build. However, there is now somereason to believe that a digital computer can beprogrammed to perform both inductively and

deductively, where its normal use in computation andcontrol is deductive, so the model we need of aphysical system may be in the form of an analoguecomputer. As a result, in automation it is perhapsdigital computers and digital methods that moreand more offer the best control facilities, and withinductive programming there seems no limit to whatcan be computer-controlled; and the systems con-trolled will involve those other stand-bys, servo-systems and their chief realisation in the form ofanalogue computers.

Theoretical consequences of cyberneticsWe should perhaps mention, before we get to the

real problems of practical interest, a few generalpoints about the scientific and theoretical backgroundof this work on cybernetics and operational research.

In the" first place, mathematicians and otherscientists have spent a great deal of their time studyinglanguages and games. The reason for this is simple.In language we find the usual communicationproblems which play a vital part in every aspectof our lives. In games we find characteristic socialsituations and problems arising, and they, of course,depend upon language. We must learn the rules ofthe games, and then the tactics, and this is exactlythe way in which we organise any enterprise what-ever, and should not be regarded as an amusingbut irrelevant study by mathematicians who livemany floors up the Ivory Tower.

There has been developed a branch of mathematics— very much a branch of probability, like Informa-tion Theory—wherein games of chance are analysed,with respect to, say, our economic behaviour, althoughagain the model used can be applied, or extended,to cover any sort of situation whatever. From thispoint of view science can be regarded as a gameagainst nature, and various optimal tactics have beenworked out for the playing of such a game. Thesewe would automatically use whenever we started toapply operational research, or our learning pro-grammes, to actual industrial situations. Similarly,science can be regarded from the point of view oftranslation of the coded language of nature, wherewe seek to discover the rules of the language andthen apply the rules deductively. These two applica-tions are of fundamental importance to cyberneticsand operational research.

We shall mention an interesting result derivedfrom mathematical logic which shows that any game— such as chess or draughts or noughts-and-crosses —if it must be completed in a finite number of moves,these must be capable of having a deductive pro-cedure. Logicians have called this a decisionprocedure; which means that there is a book of rulesthat says exactly how you should proceed to play thegame and never lose. Games are not limited ininterest since we can regard almost anything as agame, and we know now that if it is also finite it canbe " machine played" — since this is surely thecorrect interpretation to place on the idea of playingfrom a book of rules.

If we accept this criterion and agree that a deduc-tive process is certainly one that we might expect to

636

reproduce in a machinelike process, we shall notmean to imply that those processes that are notcapable of being so mirrored are not capable of sucha machine interpretation.

Indeed, the history of automatic control systemsand their application, suggests that we normally tryand reduce what is primarily an inductive process —something for which no rule book exists — to adeductive process, so that we only have to build adeductive system to carry out the necessary operation.This does not, of course, bring out the other featureof the operation, which is that, however complex theprocess, it must be capable of being reduced to aseries of simple steps. This is, in fact, a necessarycondition both for automatic control systems andalso for the ability to perform automatic computa-tions as they are carried out by the computer.

Cybernetics will go ahead and define all systems— whether organic or otherwise — in terms of themodels and theories which can be supplied for itspredictive description, and from this point of view,it is claimed that living and non-living systems arenot essentially different. They both involve com-munication and control; where to control a system,the possibility of communicating in the form ofobserving the existing state of affairs, and sendingout a controlling message to change (if necessary)that state of affairs. Here we tend to emphasise thatcybernetics is capable of supplying predictive modelsfor the biological sciences, and we can leave it at that,except to add that much time and effort has beenspent in this direction in building models andtheories which are effective in reproducing every sortof empirical situation.

Our last word on the theory of the subject is to theeffect that certain specific techniques have provedvery useful in our process of scientific model-making.We might mention such developments as informationtheory —• which is the general theory of communica-tion between any two points whatever. Also stochasticprocesses, and particularly Markoff processes, whichare the forms of statistical analysis that are mostclosely related to the description of empirical situa-tions involving human and other inductive systems.Tn this context we might mention mathematicsgenerally and mathematical logic in particular, andperhaps Theory of Games and abstract algebra.

Stochastic modelsBriefly, let us show a stochastic model and its

importance. Consider a language which has onlythree symbols and here we choose such a languagefor the sake of simplicity. We could, as does thecomputer, use a language with only two symbols,or English which has 27. allowing one punctuationsymbol.

If our alphabet is made up of just three symbolsA. B and C, say, then any sequence of symbols:

ABBACCBBABCBACAA

is a stochastic process, and we may associate aprobability with each letter. If each has a probabilityof £, then the three letters will tend to occurequally often.

Now a Markoff process is a stochastic processwhere the probability of occurrence of one letter isconditional on the probability of the letter thatoccurred immediately before. So for the table :

ABC

A

1

i

Bi0i

ci0i

we may expect, starting with A, to find sequenceslike:

AAABAABAAACAACCBABAAAC

where the occurrence of the letter is dependent onthe letter immediately before. If the source emittingthese letters is said to be ergodic, then we shall expecta degree of homogeneity and series of letters will tendto repeat themselves. Indeed, we can build machinerythat operates on these principles. Furthermore, weshould regard a machine that operated in suchMarkoff process terms, where the probabilities in thetable changed with experience (by keeping a count)as a learning machine, and exactly what our learningprogramme aims to copy.

Machine translation of language obviously fits inwith these sort of models, and science generally fitsin if we regard nature as an ergodic source.

This aspect of our subject links up with informa-tion theory and whole ramifications which we haveno time to discuss, but whose applications are onlyjust beginning.

This must suffice for a summary of the vasttheoretical background to our problems, now to theirmost general practical application and some of theirconsequences.

Science in action

Perhaps we can divide the practical applicationsinto three categories :-

1. the engineering category with all the attendantdifficulties of planning a factory at the floorlevel;

2. the problem of organisation at all the variouslevels at which it occurs — this is probably themost important of all current cybernetic andoperational research applications;

3. the industrial psychological problem involvinghuman relations ; can we expect to createcircumstances where a worker will get as muchpleasure from doing his job as he does fromplaying his games, and if so how ?

The first question is that of equipment. Our modernfactory installs machines to improve the efficiencyof the muscular output of man. and this it has donefor years. It results in increased production, andincreased efficiency in the sense of an article producedmore economically.

As far as the equipment problem is concerned, it isfairly well-known that one of our biggest currentproblems is in fitting the appropriate control to themechanical system that is to be controlled. Thisproblem has been brought into high relief by thefact that mechanical and control engineering have

637

largely developed on separate lines. The problem offinding the suitable machine to control is largely aproblem for engineers, and our only comment onthis will be to the effect that we must remember toanalyse exactly what it is that we are hoping toproduce or achieve, and to remember that we maybe able to achieve the desired end with a whollydifferent form of apparatus from that which we haveconventionally used. This problem has two sorts ofconsequences to which we should draw attention.

We may undertake to build an automatic controlsystem to replace a human operator pilot; forexample, we might proceed to build a system thatsimply does the work of the human pilot in his con-text as a pilot, therefore omitting many otherabilities that he may have which do not bear on hisflying skill. But further consideration will persuadeus that this is where we need to rethink — in thewords of John Diebold — the problem; it would, infact, be grossly uneconomical to approach it bydirect substitution of control. In the first place, the.layout of the cockpit of an aircraft is explicitlydesigned around the human operator, with dials andpointer readings that appeal to the human eye, andcontrols that are designed for human arms and legs.These dials and controls are only specified ways ofobserving and changing the state of certain systems,such as the engines and their speed, and variousengine states. This suggests that we shall, for ournew automatic control, feed information about thedifferent systems directly into control and make ourmodifications by linkages which are suitable to ourservos, or computers controlling the changes.

One important point emerges from this and that isthat we have come to regard certain systems as beingconveniently observed in terms of certaincharacteristics that lend themselves to easyrepresentation on dials, whereas we might now con-sider whether we cannot discover new and perhapsbetter indices of the states of our various systems.The pilot's eye scanning the surface of an aircraft willdiscover a fire after it has broken out; could we notfind some index of the occurrence in the future ofcombustion and make direct use of such an index(such work has in fact already been carried out) ?All this implies the development of new technologicalmethods that will accompany, and even makepossible, the machinery of the future, where suchmachinery will be designed in terms of the controlsystems on which it will operate.

As for control, we see already that digital methods,so successful in mathematical computations, havebeen developed to deal with mechanical outputs aswell as mathematical ones. We see this emphasisedin many new developments; among those best knownare the Ferranti and EMI computer-controlledcutting machines. Analogue systems have, of course,also played a large part in such control.

In terms of what we said earlier about inductiveand deductive systems, it seems that existing digitalcontrol methods are normally deductive in their use.We should lay heavy emphasis on the fact that thisdoes not need to be the case, since a computer is nota fully defined machine without its programme, and

experiments are now going forward on the use ofinductive programmes which should prove of theutmost value in the future. The idea underlying thisdevelopment can be described briefly.

ComputersA computer normally follows explicitly the instruc-

tions that are put into its storage and theseinstructions are followed regardless of what particularnumbers occur in the storage. With an inductivesystem, we shall let the instructions vary accordingto the nature of the inputs that subsequently occur.To draw the analogy with ordinary human behaviourwe should want to say that the input was cominginto the system all the time. In practice, we shall usethe computer inductively wherever this situation isunavoidable. This reminds us that the uncertaintyoccurs under these circumstances precisely becausewe do not know in advance what is to occur in theinput. We can reduce this case to the more unnaturalcase where the numbers may be stored and thenoperated upon in some order, and the instructionsmodified in the light of the items on which theinstructions operate.

More specifically, we shall have in our programmeto specify certain designated input letters, so thatthe occurrence of these designated inputs serve toreinforce or inhibit the combination of inputs that itfollows.

Then for the instructions to modify themselves inthe light of " experience", we must cater for theclassification of information in storage so that theinductive and deductive consequences, as well asother patterns and relations bearing on what isstored, can lead to new generalisations (i.e.instructions).

The detailed problems of the inductive machine,and actual programmes that have been developed, bymyself and my colleagues at Bristol, and similar pro-grammes, particularly on many-valued logicalprogrammes, by Professor Porter and his associatesat the Imperial College in London, will lead to newspecial purpose computers, new programmes forgeneral purpose computers, as well as the develop-ment of new digital, analogue and digital-analoguemethods.

We should notice in passing that this is a matterthat is of as much interest to the biologist and thepsychologist as to the production engineer, strivingas he is to fight the economic problems of inflation,by the systematic applications of the unlimitedresources of scientific endeavour.

Logical netsI wish now to outline a special set of models which

I have myself played some part in developing, andwhich I believe may be of special use, ultimately insupplying a universal language in which to constructverbal models; because we must remember thatmodels are not to be thought of only in hardwarebut also as abstract sets of symbols — indeed, any-thing at all that can be given an interpretation ormade isomorphic with a system that we wish to

638

describe or study. From this point of view, we would,for example, distinguish between marks on papersuch as 0, 1, 2 , . . . , + . —, and so on, and theinterpretation that we place on these symbols as thefamiliar operations of arithmetic, and we call them"zero", "one" , " two" , . . ' . , "plus" , "minus"and so on accordingly. Normally we do not botherto make this distinction between the concept and thename of some symbols which refers to the concept,but it is this distinction that is necessary to the preciseuse of models and language — which we couldregard as models for our thought — in science.

VVe have now developed such a model for allpossible computers and we should state straight awaythat in this respect it is slightly different from anearlier method for all computers by Turing. Turingconstructed a theoretical machine which has sincebeen called a Turing Machine and which was madeup of a potentially infinite tape marked off intosquares, with symbols So, Sx, . . . , Sn that are writtenon the tape, and the scanning and writingdevice that makes up the rest of the machine.It is capable of overprinting, moving to the leftone square, to the right one square or stopping.The machine can be said to be in anynumber of different states, for which we use thesymbols q1} q_>, . . . , qm and then a particularTuring Machine is defined by a set of quadruples,which are four symbols such as Siqk—>Sk and whichsays " when the machine is in state qk scanningsymbols Si it will move one step to the right on thetape and scan symbol Sk which is the next symbolprinted on the tape ".

With only four different such instructions andsome special conventions for the input and outputof numbers, such a theoretical computer can dealwith any problem that can be handled by the fulldeductive use of any computer.

It was in terms of the Turing Machine that Turinghimself was able to show the class of mathematicalproblems that were solvable or computable by themost obvious of " machinelike" methods. Perhapsthe most curious thing about his machine was the factthat he was able to show that it had great generality,and that any other machine that could be designedfor such computations was capable of being reducedto a computing machine of the Turing type. This isvery much the same situation as with models of com-putations in terms of many-valued functions. It canbe shown that there is no loss of generality in con-sidering only binary functions, since all problemsinvolving more than binary functions can be shownto be reducible to the binary form.

Turing Machines can be called infinite orpotentially infinite automata. I will now turn to ourmore central case of finite automata. A particularform of finite automaton — another theoreticalmachine — can be described as follows.

The elements of the system are the basic buildingblocks and these elements are connected to eachother by lines that are divided up into input andoutput lines, where by convention we draw our dia-grams showing elements as circles and input lines onthe left and output lines on the right. There are two

different sorts of possible input lines which we callexcitatory and inhibitory, and these tend to exciteor stimulate on one hand, and inhibit the elementon the other. There is only one output line whichmay divide into a number of different lines going toany number of different elements, but all must be inthe same state at any instant: either carrying animpulse or not carrying an impulse. Now eachelement has a number associated with it and thismeasures the sensitivity of the element and says justwhat the preponderance of live excitatory inputs mustbe over live inhibitory inputs to fire the element inquestion.

Now we can divide our elements into threedifferent classes, those that occur, conventionally, onthe left of our diagram and without elements attachedto their input wires — called input elements; thoseat their right-hand end and without elements attachedto the output wires and called output elements; andthe remainder which we call inner elements. Thesesets of elements taken together we call a network ornet and it can, of course, be regarded as a model of acomputer. We have described our computer in binaryterms, without, as we have already said, any loss ofgenerality.

With these theoretical computers we can investigatethe properties that any computer, or any controlsystem whatever, may have. We can show, forexample, the process of input, output, storage andarithmetical operations that make up the ordinarydigital computer.

We can show simple properties of association in anet where messages can become associated with othermessages in such a way that a message originallyhaving a particular effect, now has a different sortof effect, or a new message takes on the same effectas an older message.

In short, we can reconstruct all the characteristicsof deductive machines and envisage principles suchthat any type of control can be investigated. How-ever, this is not the main interest that attaches tosuch work. The main interest is that we can now goon and show that inductive machines can be con-structed, so that general principles can be inferred asa result of conditional probabilities from a count ofparticular instances. This can be achieved in avariety of different ways, and we can realise in ourmachine any probability function whatever. This hasnaturally led to the realisation of certain statisticalactivities such as those represented by Markoff pro-cesses. This is the situation, already mentioned, whereone event is followed by another with a certainprobability. We can construct a computer which willoperate on such a control principle, where theprobabilities can be shown to change as the resultof the counting operation of the machine.

The power and effectiveness of this sort of model-construction process lies partly in the fact that wecan define precisely each step in the construction ofthe machine by a mathematical logical formula. Thismeans that there is a relation between our diagramsor drawings — real or conceptual — and a certainpart of mathematical logic, and we know that forthat part of mathematical logic there is an effective

639

process ^sometimes called a decision procedure) whichlets us know that anything that can be constructedwithin the confines of the logic can be constructed,m fact, however many elements are involved andhowever complex the ultimate relations. We shouldnotice carefully that this means that we can drawup a blueprint, and suggest a method by which apaper machine may be constructed without giving anelement-by-element description, but by the use of ageneral formula we can describe a machine that weknow can be translated, if necessary, into a hardwaremachine.

We have at Bristol built two of what we hope willbe a series of computers to realise some of theprinciples which are implicit in the paper machinedesigns. These computers are themselves only steppingstones towards bigger and more interesting computersthat will illustrate the inductive control which wehave already been able to demonstrate in the theory.This sort of work parallels that of programmingexisting digital computers to learn, and these twotechniques taken together must be seen as part of theeffort to find appropriate control systems tocontrol effectively any process whatever. Preciselythe same features of motivation and storage haveto be built into special purpose computers (boththeoretical and hardware) as into the programmedgeneral purpose computer.

We should also mention, that our work in this fieldis by no means isolated; there are various effortsbeing made along the same lines, and the names ofDrs. Grey Walter, Ross Ashby, A. M. Uttley, D. M.Mackay, Colin Cherry and many others besides,spring to mind as having made such models — orspecial purpose computers — and broadly speakingtheir motive has been the same as that stated here.

Implications of these new ideasIt was said recently by an American industrialist

that nothing was so certain to bring economic disasterto a firm as a good scientific idea in the hands of arival organisation. This same state of affairs is trueof countries as much as it is of firms. At the moment,the scientific ideas underlying cybernetics andoperational research represent some of the mostpowerful of all ideas for production.

The methods of computation we have referred toare as important—if not more so from the engineeringpoint of view—in the control of mechanical opera-tions as in the role of a standard computer, but ournext set of problems places a different emphasis onthe whole matter.

In the immediate past the bulk of the efforts onautomatic control systems (collectively calledautomation) has been located in engineering proper,and has depended on servosystems and other fairly-simple closed-loop systems. This is all to the good, andlatterly it has been increasingly recognised that com-puters could automate the office, with automaticfiling systems, accountancy machines, and so on. Butstill insufficient emphasis has been placed on thebroader organisational aspects of automation. Thisis the field where operational research and cyberneticsmeet.

We have already said that operational research isconcerned with applying scientific method to anyproblem that does not come within the scope of theestablished techniques of science — a problem that isnot reducible to a known type of problem in physicsor chemistry, for instance.

The methods are the methods of systematiccommon sense, aided by the concept of science as amethodology wherein we may separate the predictivetheory from the model which underlies the theory.This construction of a model, therefore, whilecertainly an analogue process, is the basis of thesubsequent predictive theory which can be derivedand from which we can deductively order ourproblems, however large. The models which are usedhere will not necessarily be physical — in hardware— but in symbols, they will be logical models, and inparticular, they will frequently be mathematical,often mathematical-statistical. A set of such modelconstruction methods has been elaborated by thewriter for use in almost any situation. This set ofmodels was called the logic of empirical relations,and involved the mapping of the classical calculi ofpropositions, functions, relations and so on, on tothe calculus of probability. This means that muchthat is now discussed in ordinary language is capableof more precise restatement in our model language,and thus by suitable choice of models to any degreeof precision whatever. This at first sight seems tomake problems intensely complicated and, indeed, inmany cases it may. The important thing is that ifwe have complete precision of language the com-plication will no longer matter, since all our modelscan be put in a form where they can be handledby a digital computer.

Let us now consider some of the applications thisentails. We shall be hearing later of certain applica-tions from Mr. Harling, employing techniques notsubstantially different from these we are describing,and within the scope of operational research. How-ever, we may mention some problems where ourtechniques might be expected to yield dividends,first of all, in the organisation of any department ofa large firm and, secondly, in the fitting together ofthe departments of these large firms, to make foroverall organisation.

It is well-known from existing work study experi-ments that a great deal of time is lost in operations ofa sequential kind that contribute towards a totalproduction line. This problem is partly a matter ofcommunication theory, and the application of themathematical theory of communication in such acontext is strictly a piece of operational research.This has actually been done in certain organisations,where efforts have been made to trace all theconnections in the information network, to seewhether or not the information is adequate, suitablycoded and properly timed. The lack of these facilitieshas been shown to be a common cause of organisa-tional inefficiency.

Apart from information flow we might model allthe neural-muscular responses made by all the humanoperators, or the operations of machines that arelinearly connected in a production outline. We have

640

clearly to maximise the time flow, which might beslowed — as the ships in a convoy — by the slowestoperation.

But apart altogether from the problems oforganisation for the maximum physical efficiencythrough effort, there are problems of a more obviouslytechnical kind like those of production schedulingprogrammes, linear programming for economicproblems, and so on. There is some hope that theseare matters that might be handled by the con-struction of a suitable model and the use of a suitablecomputer programme.

I would like now to turn to the third and perhapsmost promising of all the possible developments, alsoI think within the domain of operational research,but also involving the use of computing machinery.

We know that what we commonly call " goodwill "is more important than quite large differences inphysical conditions that might pertain at anyparticular time. Industrial psychology, andparticularly experiments such as the Hawthorneexperiment in America, has shown the importance offactors of human relations, and we must remind our-selves quickly of these essential facts. There is acertain range of physical variables that will directlyaffect any human organisation, such as a factory.For example, the lighting, the hours of work, thesalary, the various conveniences and social facilities,the holidays, all certainly make a difference towhether one accepts a job in a particular firm at aparticular place. But overriding all these considera-tions to a surprising extent are the psychologicalvariables. Let us quote a commonplace example.

Recently there has been a great deal of argumentabout psychotherapies. Is it a fact that psycho-analysis improves the state of those who are exposedto it ? Indeed, are any psychotherapies effective in" curing" those people who we may all agree aresick — " mentally sick " ? The answer is to try anexperiment, and find a large group of people whohave certain complaints, and divide them up in sucha way that we have little groups that all show thesame symptoms to about the same degree. Now usethe various therapies on all but one group and leavethat one entirely alone. Now compare results aftersome period of time. This experiment has beentried under somewhat limited circumstances, but theresult seems to suggest that you will have at leastas a good chance of recovery if nothing is doneto you, as if you receive some form of therapeutictreatment.

There is an obvious " twist" to the facts statedabove. It is not possible to find the patients withapproximately equivalent symptoms without carryingout a fairly careful analysis of all the people con-cerned, and it seems as if the mere fact of carrying-out such an analysis improves many such cases.

The Hawthorne experiment seemed to demonstrateexactly the same fact, that it is not — within certainlimits — what you actually do, but it is the mannerin which you do it that matters most to most people.It seems that affection and an interest being takenis enough to give almost every one of us a consider-

able fillip, and make us exhibit the spirit of socialco-operation necessary to any enterprise.

This all seems easy enough, since it should not beimpossible, once an employer has learned these facts,to apply them by ensuring an interest is taken in hisemployees; he may ensure that the contact betweenthe different levels of his hierarchy of organisationis good, that the employer-employee relationship iseverywhere personal. But this is not easy, due to thehistorical background of suspicion between employerand employee, because of the economic pressure thatmay leave very little scope for freedom of thoughtand action within the firm. However, this is certainlya matter that can be helped by operational researchand cybernetics.

If a firm gets too large and unwieldy, it tends tobecome impersonal. If we can find appropriate indicesof impersonality, then we can ensure — subject tothe engineering and transport requirement -— thatthe firm is split up into smaller units and a suitablepersonal size is retained. But this example is only oneof a host of examples that can be treatedscientifically, where we mean the word "scientifically"in its very broadest sense to include all the psycho-logical variables as well as the historical backgroundin which a problem should always be viewed; this is acharacteristic of operational research.

The process of operational research we havealready described as involving the construction ofmodels, and these final social and industrial modelsinvolve in the limit nothing less than whole collectionsof individuals from various backgrounds workingtogether as a unit.

We need to build a social-industrial model, whichwill suggest principles of selection and organisationfor any firm. The size of the ultimate plan and theamount of detail needed is awe-inspiring and lookswholly incapable of being handled, and the answerto this is that under ordinary circumstances it wouldindeed be quite impossible to handle, but it ispossible as long as we have the use of a digitalcomputer.

The problems of learning machines and learningprogrammes, the problems of linear programming,the problems of the scheduling of organisation atevery level of the largest firms, are all particular casesof the more general problems that involve nothingless than the total organisation of any community orgroup.

To make this possible we have to designateindividuals by more than just a single-dimensionalvariable. We must decide — and this demands astudy of our learning programmes — exactly whatis the minimum set of variables that will designateany set of humans, and will allow for the predictableinteraction in specified conditions. This way we canarrive at social indices, within which we can try andmaximise the happiness, as well as the output, of thegroup as a whole. A study of such a system shouldalso reveal the sort of person who should be selectedfor employment, those most suited to promotion, andso on. Not that it is intended to replace direct humanobservations and human considerations, but rather tofacilitate them.

641

Ultimately the number of human operators will bereplaced by machines, and again we might expect ourprogramming to facilitate the changeover fromhumans to machines, by studying the human in avariety of different and characteristic situations.When the change is largely effected, this still leavesour social computer programmes as applying directlyto the machine control systems themselves, since ourcybernetic theory makes no essential distinctionbetween machines and organisms.

Programmes have already been constructed whichaim to model social situations, stochastic models havebeen used to characterise the human being in socialsituations and these models could be used as the basisfor a computer programme. Theories of learning,which are essentially the same, and theory of gamescan be also put in a computer programme form.

From these more general programmes, can beextracted models which will mirror virtually any sortof organisational situation with, of course, thepractical problem of eliminating wastage due to badselection, and bad internal organisation, both inordinary physical terms and also in psychologicalterms.

Perhaps it should be stated explicitly at this pointthat the benefit of these operational research tech-niques are more in the attitudes they engender andthe concepts they produce than in their immediateand particular application. Queueing Theory, settheory, recursive function theory, and all the otherpossible fields for models that we have previouslyreferred to, are abstract models that can be usedto represent different physical events by analoguetechniques. But this use of models is only effectivein the general background of rigorous scientificapplication.

At the machine level, we shall expect the applica-tion of general theories, such as those which logicalnets should produce as particular cases, the methodsby which we may construct self-maintainingmachines and self-servicing machines, and machinesfor quality control. In general we might expect tomodel anything whatever, even to the extent thatmarket research activities, and other sales activitiesand research, could be modelled and thus carried onwholly by machine.

These are all general statements which can beproved at the theory level, but mostly need to bedemonstrated at the practical level. Needless to saytheir application to particular problems may involveconsiderable practical difficulties, not the least ofwhich will be that they are uneconomic to apply, ormay be at any given time.

Practical approachesIf we are to make use of the very powerful

weapons with which science is able to supply us, wemust have practical experience of the situations towhich they are applied, as well as the scientific know-ledge to apply.

The biggest barrier to this application in the U.K.has certainly been in the relatively backward roleof the technical college, and the inadequacies of our

educational system, especially in technical education,where there is a tendency to teach particularitiesrather than to give the general scientific trainingwhich is essential if we are to be able to makeapplications. This implies the urgent need toreconsider our educational system, and for industriesto take very seriously the educational services theyoffer. Generality should be much more the aim in thefuture than it has been in the past.

Apart from the various problems that the educa-tional process throws up, there are the practicalapplications of all we have said. How do we setabout applying such knowledge ? The answers areso various that to generalise is now difficult, butwe can say that the first step, granted the necessaryknowledge of science, is to investigate any firm'sorganisation from the broadest possible point of view,with the idea of constructing detailed flow chartsfor all the processes carried on, with a clear picture ofwhere all the external inputs and outputs come from(and go to) and what this entails both economicallyand in engineering or other organisational terms. Itis not reasonable to analyse and automate — inthe widest sense we have been implying — any partof a system without a fairly good knowledge of thatlarger system, of which it is a part.

From this picture we can abstract that part whichlends itself most readily to replacement of humanoperators by automatic control systems, and thisdepends on what our control systems are capable ofat any given time. Then this may also mean, takingautomation again in its widest sense, the considerationof communication and other organisational problemswithin the department, or the firm, irrespective ofwhether or not automatic controls are to replacehuman operators.

The larger picture is vital to this process, becausethere is little point in greatly improving theorganisation of a section that is already one of thefaster sections in a production line.

Also we have to consider the effect on the rest ofthe firm if some part of it is made automatic, andnot the least of these considerations is the effect onthe human beings who are involved, whether or notthey are displaced.

This also suggests that we should always considertwo vital factors; what changes are likely to occurin the foreseeable future so that we can show themaximum amount of flexibility in planning, takingonly a certain minimum of things as fixed and makingeverything else easily changeable. And this leads tothe second point: is it not possible to rethink a partor the whole of some production operation ? Andif this is not possible all at once, then in what ordershould it be done ? Failure to rethink sufficientlyto allow careful planning can seriously hamper anapplication of automation. Where smaller and newerfirms are concerned, the planning-as-a-whole willautomatically be a vital factor, but even here it mustbe elastic and facilitate change.

In particular the first step for any department is todocument carefully exactly what is performed,specifying the order of the operations in great detail.

642

This should lead to a detailed flow charts which issomething that could be coded into a computer pro-gramme. Then the programme, which may, in thelimit, include human factors — where they areinvolved — will be examined with the simple ideaof making it optimal, with respect to certaincharacteristics, of efficiency time, etc.

These detailed analyses of operations actually per-formed have already been used to reveal methods ofan improved character that increase production. Herethough we. are thinking of them as stepping stonesto setting up careful programmes which will auto-matically compute wastage points and otherweaknesses in the production situation.

Finally, we might feel obliged to answer a possiblecriticism — and it is a very real one — to the effectthat what has been said here offers really nothingnew. Is it not already a fact that industrial psycho-logists have long worked on time-and-motion studies,economists and statisticians on linear programmes,and organisers tried to improve the facilities in anyfactory, and this most recently has been mainly inthe consideration of automatic control replacinghuman operator control ? The answer to this is thatwhile all these things have been done they have neverrepresented an integrated effort to solve the problemsof the factory or the production line all at once.It is not perhaps surprising that attempts to under-stand human behaviour, the human nervous systemand the other systems of the body have increasinglyemphasised the need for seeing the problem as awhole, and it is this method that seems essential tofollow up in the future.

Granted that we must analyse and then re-synthesise every aspect of an industry, andcontinually do this, and provided we use logical andscientific methods in the way I have been describing,

will it not be the case that the problems will be fartoo big for us to handle ? After all, attempts tosolve the production scheduling for batches, forexample, is seen to be exceedingly complicated, andnot one that a logical programme could readilyhandle, and yet this is only a small part of the biggerpicture. The answer to this is that we are certainlysuggesting the systematic and wholesale use ofscientific models, we are certainly suggesting thepossibilities of new programming techniques in theapplication of computers, where digital computers areincreasingly being used as simulators of various com-plex systems. The trick is to partition our problemsfrom our large scale scheme of things in such amanner that although all the individual problemsare stated in a separate handleable form, they arenevertheless structurally related to all the otherproblems that lend themselves to a computer pro-gramme. This means that we shall not be consideringone set of computer results independently of othercomputer results; this really implies the use of acommon language and, therefore, a common model,for all our organisational problems, and one aspect ofthis is the use of automatic machinery to do jobspreviously done by human beings.

To supply these models is the next step, and herewe have mentioned the various models that have sofar been worked on, often with quite different under-lying motives. Logical nets, stochastic and otherstatistical models, and computer programmes, bothdeductive and inductive, might be mentioned here;what unifies them all is the theory-model approachto scientific method and it is the application of this,as much in changing our attitude towards problems ofproduction, as anything else, that represents the bigpossibility inherent in cybernetics and operationalresearch.

" SELLING IN WORLD MARKETS " — concluded from page 609The Chairman pointed out that a lot of people were

scared of the problems of exxhange and so on. They hadbeen so busy dealing with their home market that they hadnot been prepared to face difficulties, some of which wereonly in their own imagination.

The President (Lord Halsbury) said that for a smallcountry which might be exporting only a small proportionof its product, the total number of people required to make asuccess of it was relatively small. The U.K. exported 20%of its manufactures on an average, and up to 40% in someclasses. Were 20% -40% of our best people assigned to thejob ? If not, was it because they couldn't be spared ? Hedid not know the answer to such questions.

Dr. Hague said in some ways there could be quite adifference between sending British equipment to, say, BritishBorneo and Venezuela. A background of usage of Britishmaterials •— maybe with little or no American competition—• was on easier export ground than where the Americanshad the " inside track ". In Venezuela, for example, Britishand American oil companies were working alongside oneanother — comparisons were naturally keen.

Mr. Hughes said he was very much struck with the rateof increase in exports to the United States in recentyears. It had gone up by leaps and bounds. This mightjustly be attributed in some measure, he thought, to thefact that for 10 years now there had been a special effort tosend exports to the dollar countries, particularly the UnitedStates. People had really taken a lot of trouble to developtheir exports to that market.

Dr. Hague asked Mr. Hughes if he could say whatpercentage of this upswing in U.K. exports could beattributed to the automotive industry — he thought it mustbe quite large. He went on to say he had been pleasantlysurprised at the success British cars had had in recent yearsin the U.S.A.

Mr. Hughes said he did not remember the exact figuresbut cars had certainly come up enormously in the last yearor two.

The Chairman said he was surprised more people didnot go to the dollar countries referred to by Sir JohnTaylor, the countries of Latin America. They were so mucheasier to get into than the United States and they were notonly extremely friendly, but very profitable. It wasunfortunate that they had been neglected.

The Chairman, in bringing the meeting to a close,expressed the view that the Forum had been very satisfactory.Clearly there was great interest in the subject and thequestions had been of an extremely high quality.

He thanked the members of the Forum and expressed thehope that they would not feel their time had been wasted.What they had said would be read by many members of theInstitution and other people in due course and would,therefore, reach a very wide audience.

The President (Lord Halsbury) thanked *he Chairmanon behalf of the Conference for presiding with so muchgrace, charm arid affability over the proceedings.

643