The Sciences of the Artificial (The MIT Press)

by Herbert A. Simon

There are two related ways in which simulation can provide new knowledge—one of them obvious, the other perhaps a bit subtle. The obvious point is that, even when we have correct premises, it may be very difficult to discover what they imply.

Thus, a NASA-launched satellite is surely an artificial object, but we usually do not think of it as “simulating” the moon or a planet. It simply obeys the same laws of physics, which relate only to its inertial and gravitational mass, abstracted from most of its other properties. It is a moon.

Under these circumstances the main route open to the development and improvement of time-sharing systems was to build them and see how they behaved. And this is what was done. They were built, modified, and improved in successive stages. Perhaps theory could have anticipated these experiments and made them unnecessary. In fact it didn’t, and I don’t know anyone intimately acquainted with these exceedingly complex systems who has very specific ideas as to how it might have done so. To understand them, the systems had to be constructed, and their behavior observed.

link to the idea of a startup identifying a need by building something and seeing if users like it

Uncertainty calls for flexibility, but markets do not always provide the greatest flexibility in the face of uncertainty. All depends on the sources of the uncertainty. If what is uncertain is a multitude of facts about individual and separate markets, then decentralized pricing will appear attractive; if the uncertainty encompasses major events that will affect many parts of the organization in the same direction, then it may be advantageous to centralize the making of assumptions about the future and to require the decentralized units to use these assumptions in their decisions.

Uncertainty is especially troublesome when it involves expectations by one unit about what other units in the same organization will do. Left to the market, this kind of uncertainty leads directly to the dilemmas of rationality that we described earlier in terms of game theory and rational expectations. Absorption of the uncertainty by the organization through managerial coordination may be the most effective course. We see in uncertainty a frequent source of advantage of organizations over markets as decision-making mechanisms.

Affected by their organizational identifications, members frequently pursue organizational goals at the expense of their own interests—that is to say, behave in a way that is altruistic from a personal standpoint. No organization could survive that elicited only behavior for which employees felt selfishly rewarded and that supervisors could enforce. The added effort that is elicited by identification is a major and essential source of organizational effectiveness and is a principal reason for carrying out economic activities in organizations rather than markets.

We can summarize our account of the respective roles of markets and organizations in a modern society as follows: (1) organizations find their niches wherever constellations of interdependent activities are best carried out in coordinated fashion in order to remove the need for individuals’ outguessing each other; (2) the human motivation that makes organizations viable and alleviates the public goods problems that arise when individual efforts cannot be tied closely to individual rewards is provided by organizational loyalty and identification; (3) in both organizations and markets, the bounds on human rationality are addressed by arranging decisions so that the steps in decision making can depend largely on information that is locally available to individuals.

An evolutionary theory of the firm might argue that it does not matter whether people maximize or satisfice, for in a world of competitive markets only those who make decisions as if they were maximizing will survive.

The EPAM theory gives us a basis for understanding what a “chunk” is. A chunk is a maximal familiar substructure of the stimulus. Thus a nonsense syllable like “QUV” consists of the chunks “Q,” “U,” “V”; but the word “CAT” consists of a single chunk, since it is a highly familiar unit.

EPAM postulates constancy in the time required to fixate a chunk. Empirically the constant appears to be about eight seconds per chunk, or perhaps a little more.

If asked to read a string of digits or letters and simply to repeat them back, a subject can generally perform correctly on strings up to seven or even ten items in length. If almost any other task, however simple, is interposed between the subject’s hearing the items and repeating them, the number retained drops to two. From their familiarity in daily life we could dub these numbers the “telephone directory constants.” We can generally retain seven numbers from directory to phone if we are not interrupted in any way—not even by our own thoughts.

When a domain reaches a point where the knowledge for skillful professional practice cannot be acquired in a decade, more or less, then several adaptive developments are likely to occur. Specialization will usually increase (as it has, for example, in medicine), and practitioners will make increasing use of books and other external reference aids in their work.

The artificial world is centered precisely on this interface between the inner and outer environments; it is concerned with attaining goals by adapting the former to the latter. The proper study of those who are concerned with the artificial is the way in which that adaptation of means to environments is brought about—and central to that is the process of design itself.

Design, on the other hand, is concerned with how things ought to be, with devising artifacts to attain goals.

it is more useful to think of the values as estimates of the gain to be expected from further search along the path than to think of them as “values” in any more direct sense. For example, it may be desirable to attach a relatively high value to a partial exploration that may lead to a very good solution but with a low probability. If the prospect fades on further exploration, only the cost of the search has been lost. The disappointing outcome need not be accepted, but an alternative path may be taken instead. Thus the scheme for attaching values to partial paths may be quite different from the evaluation function for proposed complete solutions.

Here are the rules of a game, which I shall call number scrabble. The game is played by two people with nine cards—let us say the ace through the nine of hearts. The cards are placed in a row, face up, between the two players. The players draw alternately, one at a time, selecting any one of the cards that remain in the center. The aim of the game is for a player to make up a “book,” that is, a set of exactly three cards whose spots add to 15, before his opponent can do so. The first player who makes a book wins; if all nine cards have been drawn without either player making a book, the game is a draw. What is a good strategy in this game? How would you go about finding one? If the reader has not already discovered it for himself, let me show how a change in representation will make it easy to play the game well. The magic square here, which I introduced in the third chapter, is made up of the numerals from 1 through 9. 4 9 2 3 5 7 8 1 6 14. As examples of current thinking about representation see chapters 5 (“Representing Designed Artifacts”) and 6 (“Representing Design Processes”) in C. L. Dym, op. cit., and chapter 6 (“Representation in Design”) in O¨mer Akin, op. cit. For a more general theoretical discussion, see R. E. Korf, “Toward a Model of Representational Changes,” Artificial Intelligence, 14(1980):41-78. Each row, column, or diagonal adds to 15, and every triple of these numerals that add to 15 is a row, column, or diagonal of the magic square. From ...

saw that in math museum, wonder if developed by the author?

That representation makes a difference is evident for a different reason. All mathematics exhibits in its conclusions only what is already implicit in its premises, as I mentioned in a previous chapter. Hence all mathematical derivation can be viewed simply as change in representation, making evident what was previously true but obscure. This view can be extended to all of problem solving—solving a problem simply means representing it so as to make the solution transparent.

The growing theories of computer architectures and programming languages—for example, the work on functional languages and object-oriented languages—illustrate some of the directions that a theory of representations can take.

As we draw up our curriculum in design—in the science of the artificial—to take its place by the side of natural science in the whole engineering curriculum, it includes at least the following topics: THE EVALUATION OF DESIGNS 1. Theory of evaluation: utility theory, statistical decision theory 2. Computational methods: Algorithms for choosing optimal alternatives such as linear programming computations, control theory, dynamic programming Algorithms and heuristics for choosing satisfactory alternatives 3. THE FORMAL LOGIC OF DESIGN: imperative and declarative logics THE SEARCH FOR ALTERNATIVES 4. Heuristic search: factorization and means-ends analysis 5. Allocation of resources for search 6. THEORY OF STRUCTURE AND DESIGN ORGANIZATION: hierarchic systems 7. REPRESENTATION OF DESIGN PROBLEMS

All three chapters, so construed, have import that goes beyond the professional work of the person we have called the “designer.” Many of us have been unhappy about the fragmentation of our society into two cultures. Some of us even think there are not just two cultures but a large number of cultures. If we regret that fragmentation, then we must look for a common core of knowledge that can be shared by the members of all cultures—a core that includes more significant topics than the weather, sports, automobiles, the care and feeding of children, or perhaps even politics. A common understanding of our relation to the inner and outer environments that define the space in which we live and choose can provide at least part of that significant core.

that unification requires knowledge of and acquiring of that "shared knowledge" - a tough ask from all members of all cultures

Undoubtedly there are tone-deaf engineers, just as there are mathematically ignorant composers. Few engineers and composers, whether deaf, ignorant, or not, can carry on a mutually rewarding conversation about the content of each other’s professional work. What I am suggesting is that they can carry on such a conversation about design, can begin to perceive the common creative activity in which they are both engaged, can begin to share their experiences of the creative, professional design process.

The proper study of mankind has been said to be man. But I have argued that people—or at least their intellective component—may be relatively simple, that most of the complexity of their behavior may be drawn from their environment, from their search for good designs. If I have made my case, then we can conclude that, in large part, the proper study of mankind is the science of design, not only as the professional component of a technical education but as a core discipline for every liberally educated person.

This is not an isolated example. The first generation of management information systems installed in large American companies were largely judged to have failed because their designers aimed at providing more information to managers, instead of protecting managers from irrelevant distractions of their attention.3 A design representation suitable to a world in which the scarce factor is information may be exactly the wrong one for a world in which the scarce factor is attention.

Thus a stock of inventories permits a factory to operate without concern for very short-run fluctuations in product orders. Energy storage in the tissues of a predator enables it to cope with uncertainties in the availability of prey. A modest excess of capacity in electric generating plants avoids the need for precise estimation of peak loads. Homeostatic mechanisms are especially useful for handling short-range fluctuations in the environment, hence for making short-range prediction unnecessary.

In domains where some reasonable degree of prediction is possible, a system’s adaptation to its environment can usually be improved by combining predictive control with homeostatic and feedback methods. It is well known in control theory, however, that active, feedforward control, using predictions, can throw a system into undamped oscillation unless the control responses are carefully designed to maintain stability. Because of the possible destabilizing effects of taking inaccurate predictive data too seriously, it is sometimes advantageous to omit prediction entirely, relying wholly on feedback, unless the quality of the predictions is high.8

And in saving, we count flexibility among the important attributes of the objects of our investment, because flexibility insures the value of those investments against the events that will surely occur but which we cannot predict. It will (or should) bias our investments in the direction of structures that can be shifted from one use to another, and to knowledge that is fundamental enough not soon to be outmoded—knowledge that may itself provide a basis for continuing adaptation to the changing environment.

Bounded rationality. The meaning of rationality in situations where the complexity of the environment is immensely greater than the computational powers of the adaptive system.

The fact then that many complex systems have a nearly decomposable, hierarchic structure is a major facilitating factor enabling us to understand, describe, and even “see” such systems and their parts. Or perhaps the proposition should be put the other way round. If there are important systems in the world that are complex without being hierarchic, they may to a considerable extent escape our observation and understanding. Analysis of their behavior would involve such detailed knowledge and calculation of the interactions of their elementary parts that it would be beyond our capacities of memory or computation.

I shall not try to settle which is chicken and which is egg: whether we are able to understand the world because it is hierarchic or whether it appears hierarchic because those aspects of it which are not elude our understanding and observation. I have already given some reasons for supposing that the former is at least half the truth—that evolving complexity would tend to be hierarchic—but it may not be the whole truth.

The generalization that we might expect ontogeny partially to recapitulate phylogeny in evolving systems whose descriptions are stored in a process language has applications outside the realm of biology. It can be applied as readily, for example, to the transmission of knowledge in the educational process. In most subjects, particularly in the rapidly advancing sciences, the progress from elementary to advanced courses is to a considerable extent a progress through the conceptual history of the science itself. Fortunately the recapitulation is seldom literal—any more than it is in the biological case. We do not teach the phlogiston theory in chemistry in order later to correct it. (I am not sure I could not cite examples in other subjects where we do exactly that.) But curriculum revisions that rid us of the accumulations of the past are infrequent and painful. Nor are they always desirable—partial recapitulation may, in many instances, provide the most expeditious route to advanced knowledge.