Models of my life 5

Chapter 23 Guides for Choice
A favorite question to pose to gurus is, ''What is your philosophy of life?" On rare occasions, I have been incautious enough to answer the question. A philosophy of life surely involves a set of principles. But principles for what purpose? Principles can provide a book of heuristics to guide choice at life’s branch points, a thread to keep one on the right path in the maze. Principles can also rationalize, explain, or provide excuses for choices one has already made. It’s not easy to distinguish between these two uses of principle; perhaps it is not even necessary. A philosophy of life can contain both.
In either case, it would appear to be a lot easier to have a life philosophy at age sixty-eight than at age eighteen. Or is it easier? Perhaps the process of living confuses as often as it clarifies. Perhaps one should write down a life philosophy at eighteen, before the complexities have emerged, so that one can produce it, on request, when one is sixty-eight. But that is dangerous too. One’s readers would be tempted to compare the philosophy with the life. Safer to write it at sixty-eight, or even later.
The phrase “life philosophy” sounds solemn. We must distinguish between its two different meanings. In one sense, a life philosophy is a statement of your raison d’être in the midst of your cosmic and human environment. In a second sense, your life philosophy is your picture of this cosmos, including, in center foreground, your picture of the human condition.
As to the first sense, the creature of bounded rationality that I am has no illusions of attaining a wholly correct and objective understanding of my world. But I cannot ignore that world. I must understand it as best I can, with the help of my scientific and philosophical fellows, and then must adopt a personal stance that is not outrageously incompatible with its ap-

parent conditions and constraints. I must eschew personal goals that require gravity shields or the perfection of humankind for their success.
I am an adaptive system, whose survival and success, whatever my goals, depend on maintaining a reasonably veridical picture of my environment of things and people. Since my world picture approximates reality only crudely, I cannot aspire to optimize anything; at most, I can aim at satisficing. Searching for the best can only dissipate scarce cognitive resources; the best is enemy of the good.
Already, you have learned something about my life philosophy, both the cosmological and the personal one. Let me now describe the former a little more systematically. I am a creature of the twentieth century, thoroughly immersed in its science and its empiricism. My cosmos began (probably) with a Big Bang, and has been evolving inexorably ever since through astronomical, geological, biological, and anthropological ages, the timeline magnifying gradually, perhaps exponentially, as we approach the present, and shrinking again, perhaps exponentially, as we peer into the future. Parts of the picture change from time to time, especially the parts most distant fore and aft, but not (at least in the past quarter- century) in ways that are important for a personal life philosophy.
This cosmological machine has laws, but I cannot detect in it any purpose. In this respect, also, I am a creature of my century, needing, like logical positivists and existentialists, to postulate my own goals because I cannot see that they have been given to me by any external donor. The world is vast, beautiful, and fascinating, even awe-inspiringbut impersonal. It demands nothing of me, and allows me to demand nothing of it, a little like some people’s conception, today, of a house-sharing or bed-sharing “relationship.”
But if the cosmos is indifferent to me, I need not be indifferent to the cosmos. I can seek to live in peace with it. Nor need I put the matter so negatively. The cosmos can be the source of some of my deepest pleasures. Gazing at it, outdoors at night or in a forest or through a microscope, I find inconceivable variety, pattern, and beauty, beyond the competence of human artists.
Some of the beauty of the cosmos is hidden, to be revealed only by the code-breaking activity we call science. Catching glimpses of new patterns, never before seen by the human eye, bringing them into the open, provides the scientist with his or her most moving experiences. And though we can have such experiences directly only a few times in a lifetime, we can have them vicariously as often as we wish by studying the work of our fellow scientists, present and past.
I suppose that is why I am a scientist. But why a social scientist? How

did I choose that path? To explain why (if, indeed, I know the reasons), I must return to the cosmological stage, this time the part occupied by human beings. Neither Aristotle’s “featherless biped” nor “rational animal” seems to capture it all, though the latter is closer if we place equal emphasis on both noun and adjective. We humans are minds (and consciousnesses) in bodies that move in a physical world. We are subject, without exemption, to physical and biological laws. If we fall, our bones break; if we cannot find food, we starve.
We have become the species we are through a long process of evolution. As a result, we come into the world equipped with at least some of the requisites of survival (including supporting adults who nurture us). The newborn child is ready to breathe, to suck, to defecate. It doesn’t need a life philosophy to do those things or to want to do them. It is ready, also, to learn. And whether through learning or because of the equipment it brings with it into the world, it is soon able to empathize with other members of its species: to feel their hurts as its hurts and, later, their poverty as its poverty. I don’t need to recite a full list of human traits, inborn or acquired; I have mentioned some of the more positive ones. I could equally well have mentioned human propensities for predation against our own species, and the deeply ingrained selfishness that was surely one of the prime conditions for our survival.
The human condition is often described as absurd. Surely the term is appropriate: a body shackled to a self-conscious mindor is it the mind that is shackled to the body? The wants and needs of the two parts are absurdly disparate. Can the body regard as anything but absurd the mind while it is gazing at the stars or, worse, wrapped up in its own thought? Can the mind regard the act of sex or the savoring of food as anything but absurd?
Of course only the mind, not the body, can make judgments about absurdity. And so, given the range of human needs and wants, the mind creates myths reconciling it with the body, thus turning absurdity into pleasure, beauty, and tenderness. The mind sometimes even tries to find a common denominator for all the claims of the body, of itself and of the surrounding environment. It gives this common denominator impressive names like the Good, or utility. But the notion of a single, overarching goal is an illusion. We, the mind and the body, have many needs, many desires, fortunately not all clamoring at once (I refer you back to my story, “The Apple,” and its crude picture of this symbiosis of needs and wants [see pages
180 88]). Stamping them all with the label of “utility” would be futile. The plurality is real; there is no monolithic goal.
In this committee of urges, wants, and needs, housed in body and mind,

there is no consensus about the purpose of life. Mark Twain told a story of Siamese twins who agreed upon alternating time slots during which one or the other would be in full charge. The story did not end well: Both twins had reason to regret the murder one committed while he had control. But the absurdity of the story is the human absurdity. Each of us “time-shares,” alternating our many selves. Some parts of life are spent in the enjoyment of music, others in the enjoyment of sex, yet others in the enjoyment of food, leaving lots of time for the enjoyment of mountains, the enjoyment of friends, and, for some fortunate ones of us, especially the enjoyment of science.
Of course this list is not complete; I mean it only to be illustrative. Moreover, I have left out everything except the time spent in consuming. There is work, too, and obligation and duty; a great deal of them in most of our lives. And there are sorrow and grief, which we do not count among life’s blessings, but which deepen our other experiences and give them meaning and sometimes a poignancy they might not otherwise have.
SoI am describing a human life with many goals but without a goal. Who would want it otherwise? Who would want to be free from the hundred desires that are always making exigent demands on a day that will not stretch beyond twenty-four hours? And who is capable of fashioning that master plan, that comprehensive utility function that allocates to each want precisely its proper slice of time?
Homo Rationalis
In this chapter, I have been describing my life, and also my personal life philosophy, but I have also been describing the life of Everyperson. My interest in Everyperson began in 1935 as an interest in human decision making, especially in people coping with the complexities, the uncertainties, and the goal conflicts and incommensurabilities of everyday personal and professional life.
You have seen me following that interest in these pages over more than fifty years, an interest that has never left me. I no longer feel as ignorant of the answer as I did in 1935; I and others have made considerable progress toward understanding the conflict and providing solutions. But the allocation of individual or organizational resourceshow it is done and how it ought to be doneremains a central problem of the human condition.
Pursuing the answer has led me on a long but pleasurable search through the maze of possibilities. To understand budget decisions I had to study decision making and, more generally, the processes of human thinking. To

study thinking, I had to abandon my home disciplines of political science and economics for the alien shores of psychology and, a little later, of computer science and artificial intelligence. There I have remained, except for occasional brief visits to the home islands.
At least that is one version of the story: a single-minded search that has persisted for a half-century. Perhaps it is even the true version. Another possibility is that excitement lit the path: first the excitement, after World War II, of game theory, linear programming, and the use of mathematics in economics and operation research; then the excitement of the computer, the machine that taught us how a mind could be housed in a material body.
What significance should one attach to coincidences? The demands of the problem and the excitement of the new tools lured me down the very same path of the maze. And so I was able to spend my scientific life pursuing a problem I thought central to understanding the human condition, while indulging myself in the mathematics and computer formalisms that gave me so much pleasure just in the doing. Nor was I denied the pleasures of friendship, even in professional life, for we have seen that most of my work has involved warm partnerships.
The pictures of Homo economicus and Homo cogitans that emerged from this quest have already been sketched. When, abandoning the a priorism of neoclassical economics, I looked at actual decision making and problem solving, I saw a creature of bounded rationality using heuristic search to find satisficing"good enough"courses of action. And with the help of computer simulation, my colleagues and I were able to account for the facts of human problem solving in a range of both simple and complex situations.
Economists did not flock to the banner of satisficing with its bounded rationality. These ideas still remain well outside the mainstream of economicsbut not indefinitely. For they provide a realistic picture of human choice, a picture that may instruct us about some of the puzzling problems of economics today: decision making under uncertainty, business cycles with their accompanying natural or unnatural unemployment, the role of entrepreneurship in investment, and others. But there is backbreaking empirical work ahead, for the theory of bounded rationality does not permit all one’s theorems to flow from a few a priori truths. Fixing the postulates of such a theory requires close, almost microscopic, study of how people actually behave.
Science, viewed as competition among theories, has an unmatched advantage over all other forms of intellectual competition. In the long run (no more than centuries), the winner succeeds not by superior rhetoric, not by the ability to convince or dazzle a lay audience, not by political influence, but by the support of data, facts as they are gradually and cumulatively

revealed. As long as its factual veridicality is unchallenged, one can remain calm about the future of a theory. The future of bounded rationality is wholly secure.
Homo Socialis
How do you put duty in a utility function? For a satisficing theory it’s quite easy: Simply place it among the constraints. Of course, we may also view duty as a cost we pay for society’s willingness to cooperate with us. This implies that every person has a price. Possibly so, but I prefer the satisficing view.
What duties would I impose? Starting at the weak end of the spectrum, there is general acceptance of the duty not to harm othersthe negative version of the Golden Rule. A higher, and not unreasonable, standard is the obligation to leave the world no worse off than it would have been without us. Since most people, even people in rather humble circumstances, can meet that requirement, perhaps it is proper to insist on it.
A still heavier obligation, not always acknowledged, is to leave to future generations as wide and interesting a range of options as our generation inherited from our forebears. To do so, we must accept collective responsibility for securing sustainable energy sources, preserving the environment, stabilizing world population, and somehow removing or dulling the threat of the Bomb. We have no obligation to solve all the world’s problems (there is no prospect that we could); we do have an obligation to avert irreversible catastrophe and to oppose implacably every step toward it.
When we turn to obligations to do positive good, the road seems steeper and stonier. The social scientists of my generation are Depression children, and although the Depression probably had little to do with bringing me to social science, I share the values and feelings of my generation. Given the productivity of which human societies today are technically capable, I regard the elimination of poverty (at least poverty measured against basic physiological and psychological needs) as one of the Big Goods that is actually attainable, perhaps within a couple of generations.
Distributive justice? That’s more elusive. My cosmology shows clearly that the distribution of the world’s goods owes little to virtue and a great deal to the lottery that distributes families, genes, places of birth, material resources, and other forms of access by the throw of cosmic dice. Does that call for a norm of full equality? Only if you believe that people’s aspirations must be guided by comparison with the well- being of others. That belief seems highly unproductive, as it turns the whole life of society into a zero-

sum game in which some can win only if others lose. There must be better games. If I were to select a research problem without regard to scientific feasibility, it would be that of finding out how to persuade human beings to design and play games that all can win. Clearly neither the USSR nor China has succeeded in inventing such a game; nor have we, although perhaps we have come closer.
Homo Scientificus
Does a life philosophy, in addition to the cosmological and personal, include a third element, a philosophy of science? If so, you have already been exposed to most of mine, and can learn even more about it in the Afterword that follows this chapter. If the quality of a research problem rests on the importance of the questions it addresses and the availability of ideas and techniques that hold out a promise of progress, then the study of mind is a most promising research domain. The questions it addresses have puzzled humankind since the earliest times, and underlie the most fundamental questions of epistemology, including the much-discussed mind/body problem. Moreover, understanding the nature of mind is fundamental to building viable theories of social institutions and behavior, of economics and political science. Economics dodged the problem for two centuries with its a priori assumptions of human rationality. But those assumptions are no longer fruitful; they must be replaced by a more veridical theory of the human mind.
Since the 1950s we have had the tools to study the mind. We now have a third of a century’s accumulation of evidence that the digital computer is the crucial tool we had been lacking. The computer as applied in cognitive science both provides a language for stating theories of human behavior without placing them in the Procrustean bed of real numbers and, by simulation, spins out the implications of the theories. The computer allows us to plumb mind to the level of symbols; we still wait for powerful biological tools to plumb to the neural level.
My life shows that my tribal loyalties are weak. I am a social scientist before I am an economist or a psychologistand, I hope, a human being before anything else. I believe (my third creation myth) that what brought me to the social sciences was the urge to supply rigor to a body of phenomena that sorely needed it. Physics was already too far along (I thought) for genuine adventure. The social sciences offered a field of virgin snow on which one could imprint a fresh form.

Disciplines, like nations, are a necessary evil that enable human beings of bounded rationality to simplify their goals and reduce their choices to calculable limits. But parochialism is everywhere, and the world badly needs international and interdisciplinary travelers to carry new knowledge from one enclave to another. Having spent much of my scientific life in such travel, I can offer one piece of advice to others who wish to try an itinerant existence: It is fatal to be regarded as a good economist by psychologists, and a good psychologist by political scientists.
Immediately upon landing on alien shores, you must begin to acquire the local culture, not to deny your origins but to gain the full respect of the natives. When in economics, there is no substitute for talking the language of marginal analysis and regressionseven (or especially) when your purpose is to demonstrate their limitations. When in psychology, you must be able to understand references to short-term memory and latencies and spreading activation.
The task is not onerous; after all, we acculturize new graduate students in a couple of years. Besides, it may lead you to write papers on fascinating topics that you would otherwise never have encountered. For one of the nice features of the utility function (or the committee of goals I would substitute for it) is that it can acquire ever new dimensions. Learning a new language every decade or so is a great immunizer against incipient boredom.
In describing my life, I have situated it in a labyrinth of paths that branch, in a castle of innumerable rooms. The life is in the moving through that garden or castle, experiencing surprises along the path you follow, wondering (but not too solemnly) where the other paths would have led: a heuristic search for the solution of an ill-structured problem. If there are goals, they do not so much guide the search as emerge from it. It needs no summing up beyond the living of it.

AFTERWORD
THE SCIENTIST AS PROBLEM SOLVER
Many pages of this book have been devoted to describing my scientific work and its impact. Only in a few places have I said anything about my personal style in doing science. In a recent book (Langley et al.
1987), my co-authors and 1 have said a great deal about discovery. Deepak Kulkarni and I have added yet another chapter on this theme in a paper in Cognitive Science (Kulkarni and Simon 1988), in which we concluded that the scientist is a problem solver, searching through a maze, and that the theory of discovery is a gloss on the theory of problem solving.
Scientists set themselves many different kinds of tasks: formulating significant problems, discovering interesting phenomena, finding the laws that are hidden in data, inventing new representations for phenomena and their accompanying theories, inferring the logical consequences of theories and testing them, designing experiments, finding explanatory mechanisms to account for empirical generalizations, and inventing new instruments for observation and measurement. Undoubtedly there are others.
All of these tasks use the same general kinds of problem-solving processes as do chess players in choosing moves, subjects in the laboratory in solving the Tower of Hanoi problem (see page 327), physicians in making diagnoses, computer salesmen in configuring systems for clients, architects in designing houses, and organic chemists in synthesizing new molecules.
Moreover, the “insight” required for such work as discovery turns out to be synonymous with the familiar process of recognition. Other terms commonly used in the discussion of creative workjudgment, creativity, even geniusappear either to be wholly dispensable or to be definable, as insight is, in terms of mundane and well-understood concepts.
Much of the published work about scientific discovery has until recently consisted of anecdotes, frequently autobiographical, about specific discov-

eries and their finders. If discovery requires creativity, or even genius, it would be immodest for anyone to claim that he or she had made a discovery, and futile to try to describe how it had been done.
But if discovery is plain, garden-variety problem solving, then there is no immodesty, and perhaps not even futility, in adding to the anecdotal evidence. In the next pages I will think aloud, albeit retrospectively, about some of my own scientific work, and see whether it, too, fits the problem-solving mold. In particular, I will see if I can find examples to fit each of the components of discovery that I have mentioned.
My predictions will face backward, for backward predictions are really the only ones we can wholly trust in this realm. After all, forward predictions may be influenced by the very theories we are trying to test: The theory may fit our behavior only because we have read about it and think we will do better if we follow it.
Formulating Problems
It is usually thought that before you answer a question you must state it. Or, to change the metaphor, for something to be found, something must have been lost. But is that always true? When one finds a vein of gold, was it nature who lost it? If we can find gold that we haven’t lost, perhaps we can answer questions that we haven’t asked.
We may find gold (even gold we haven’t lost) by searching for it. But that means that we have already asked, “Where can we find some gold?” But what about the gold we find when we are not looking for gold; when we are engaged in some different activity (gathering wildflowers on the mountain, for example)? At the very least, we must notice the gold; it must attract our attention, distracting us from the flowers. Do we account for this by postulating a need for gold? Or will an attention-attracting propensity of shiny yellow objects do the job? And how does the attraction of these yellow objects distract us from the flower-gathering task?
Now let’s return from gold seeking to problem seeking. Our metaphor suggests that one way to find a problem, and perhaps even its solution, is to try to solve some other problem. That doesn’t tell us where the other problem came frombut one problem at a time! We are dealing with the phenomenon of surprise. Searching for wildflowers, we are surprised to see something shining and golden in the rocks. To be surprised, we must attend to the surprising phenomenon. Hence the dictum of Pasteur: “Accidents happen to the prepared mind.”
And now we have a new problem: How does a mind become prepared?

If I am to follow the time-honored tradition of using autobiographical anecdotes as the evidence for my theory of discovery, perhaps it is time for an anecdote. My first piece of scientific work was a study of public recreation in Milwaukee (Simon 1935). A standard topic in studies of organizations is the budget process, which in this case involved the division of funds between playground maintenance, administered by one organization, and playground activity leadership, administered by another. How was this division (which was a frequent subject of dispute) arrived at?
My previous study of economics provided me with a ready hypothesis: Divide the funds so that the next dollar spent for maintenance will produce the same return as the next dollar spent for leaders’ salaries. I saw no evidence that anyone was viewing the decision in this way. Was I surprised at their ignoring accepted economic theory? Perhaps, initially, but on reflection, I didn’t see how it could be done. How were the values of better activity leadership to be weighed against the values of more attractive and better-maintained neighborhood playgrounds?
Now I had a new research problem: How do human beings reason when the conditions for rationality postulated by neoclassical economics are not met? Investigating further, I thought I could see a rather simple pattern. Those who were organizationally responsible for playground supervision wanted more money spent for leadership; those who were responsible for the physical condition of the playgrounds wanted more spent for maintenance. Generalizing, people bring decisions within reasonable bounds by identifying with the partial goals for which their own organizational units are responsible (Simon 1947, chap. 10).
Of course this is only a partial answer. It defined and labeled the phenomenon of organizational identification, a concept that has proved valuable in administrative theory. The broader questionHow do people make decisions when the conditions for the economists’ global rationality are not met (or even when they are)?remains an active frontier of research today. The central concept is bounded rationality, a label for the computational constraints on human thinking. When people don’t know how to optimize, they may very well be able to satisfice, to find good enough solutions. And good enough solutions can often be found by heuristic search through the maze of possibilities (Simon 1955a, 1982a).
Now what does this anecdote say about finding problems as a part of scientific discovery? One thing it says is that a research problem I found in 1935 has lasted me for a half century. I have never had to find another. The broad problem of accounting for human rationality has generated an endless series of subproblems: How do people solve the Tower of Hanoi problem; How do they choose chess moves; How do they make scientific

discoveries? (Newell and Simon 1972; Simon 1979a, sec. 4, 7; Langley et al. 1987).
Another lesson is that scientific discovery is incremental. An explanation for a particular act of discovery must take everything that has gone before as initial conditions. What we seek to explain is how these initial conditions lead to the next stepin this case, how my knowledge of price theory, and my boss’s desire to know how two organizations cooperated to provide recreation services in Milwaukee, led me to observe a phenomenon that initially surprised me; and how that surprise led to the concepts of identification and bounded rationality. Steps taken fifteen years later led from bounded rationality to satisficing, and from satisficing to heuristic search.
Third, the anecdote adds another to the long list of examples of surprise striking the prepared mind as a key event in discovery. What was ''prepared" about this particular mind? My training in economics, evoked in the context of a budget situation, disclosed a contradiction between what theory taught me ought to be happening and what my eyes and ears showed me was actually happening. Without the training in economics, the observed behavior would have appeared entirely “natural.” Without the observations, I could have continued in the happy illusion that the neoclassical theory of utility maximization explains human behavior in the domain of budgeting. And since my exposure to the economics profession was still rather minimal, I had not acquired the habit, so common in that profession, of ignoring the real world when it contradicts the theory.
Nothing mystical. Nothing magical. Can we simulate it? The heuristics resemble quite closely those of KEKADA, the program that Deepak Kulkarni and I used to simulate the research strategy of Hans Krebs, who found the chemical path for the in vivo synthesis of urea (Kulkarni and Simon 1988). The program experiences surprise when its expectations are not met, and reacts to its surprise by seeking explanations for the surprising phenomena. We have not yet expanded KEKADA to simulate the discovery of bounded rationality. But I might have saved myself a lot of work in 1935 had I then had KEKADA to advise me.
Laws from Data
In Scientific Discovery (Langley et al. 1987), my colleagues and I gave primary attention to inducing laws, quantitative and qualitative, from data, using programs we call BACON and DALTON, among others, to simulate the process.
Data are not the only possible starting points for inducing new laws;

theories can also be used, in conjunction with data or independently. At the limit, it may be possible to find a descriptive law directly, by deriving it from a more fundamental explanatory law. For example, Newton showed that Kepler’s Third Law of planetary motion could be derived mathematically from the law of gravitation. (But note that Newton was working backward from the law that Kepler had already discovered by data-driven search.)
Before one can find laws that fit empirical data, one must have appropriate data that look as though a smooth mathematical function could generate them. It’s the recipe for rabbit stew all over again: First catch the rabbit.
Only once in my life have I run across such data, and I cannot recall exactly when I first encountered thempossibly as early as 1936 when I read Lotka’s Elements of Physical Biology (1924). Lotka’s data show that when the number of species belonging to each genus in some order of plants or animals is counted, and the genera are then arranged according to the number of their species, the genus with the nth largest number of species will have about 1/n as many species as the genus with the largest number.
Similarly, when the frequencies with which different words appear in a book are counted, and words are then arranged in order of their frequency, the nth most frequent word will occur about 1/n times as frequently as the most frequent word. Moreover, about half of all the words that occur in a book will occur exactly once, about one-sixth exactly twice, one-twelfth three times, and so on. These relations hold for books in any alphabetic language, and the departures from regularity are small. The same regularity is seen in the populations of cities in the United States: The nth largest city is about 1/n times as large as New York. If the data are replaced by their logarithms they fall on a straight line with a slope of minus one.
What does one do with regularities like thisregularities that, at first blush, can only be described as astonishing? Sir Francis Bacon advised us to induce general laws from them, to find formulas that fit the data; and our computer program BACON (Langley et al. 1987) shows us how to do just that. Then, like the great chemist John Dalton, we should see if we can postulate a mechanism whose operation would produce the regularity described by the formula. Our program DALTON (Langley et al. 1987) simulates that process too. With BACON and DALTON we would know what to do.
I wish I could say that this was my immediate response to the Lotka data. Memory fails me. I recall my fascination with them, but not whether I pondered over them or, if I did, for how long. I do recall that when I returned to Chicago after 1942, I thought about them again. I have a clear picture of sitting in the biology library in the University of Chicago, reading

a paper referenced in Lotka’s book. I also recall talking about the matter to Allen Newell while visiting him and his wife, Noël, in their Santa Monica apartment between 1952 and 1954. But I was doing many other things during these years. The startling data on word frequencies and city sizes were not a constant preoccupation, but were more like a recurring itch that needed to be scratched occasionally.
Sometime during 1954 I found the answer. Only a few aspects of the discovery are recoverable now from memory. First, I looked for a function to fit the data. I was especially impressed by the regularity of the word frequencies at the low end of the frequency range. The simple fractions seemed to point to a formula involving ratios of integers. In fact, the simple formula, f(i) = 1/[i(i + 1)], gives the required numbers, 1/2, 1/6, 1/12, and so on. For large i, we have approximately, f(i) = 1/i2. The rank, which is simply the integral of the frequency, will then give F(i) = 1/i, so that on a logarithmic scale the relation between rank and frequency will be linear with a slope of minus one.
Finding an equation that fits these magic numbers sets a new problem: finding an explanation for the equation, a plausible rationale for the phenomena. My recollections of how I did this are even sketchier than my recollections about the previous stage. The ratios of integers were again the key. Where can you get ratios of integers? Ratios of factorials are one possible source: 1/6 can be written as the product of 1/2 and 1/3, and 1/12 is the product of 1/2, 1/3, and 2/4. In general, the formula (i-2)!/i! produces the required numbers.
The next step is likely to occur only to someone who has a little mathematical knowledge, and sees in these ratios of factorials something like the Beta function, or at least sees the kinds of expressions one encounters in problems on combinations and probabilities. (In fact, I discovered that the Beta function was what I wanted by searching through my copy of Peirce’s A Short Table of Integrals, where I vaguely remembered having seen some ratios of factorials.)
Are there any other reasons for thinking of a probability model? Indeed there are. What do word frequency distributions and city size distributions (as well as the other quite different phenomena where this same law applies) have in common? Nothing very obvious, unless they can be viewed as instantiations of the same probabilistic scheme for drawing balls of different colors from an urn. So let us see whether we can interpret the formula as representing the steady state of some sampling process.
Here I recall being aided by a metaphor. We think of a book as being created word by word. If a word is added that has already occured k times, the number of words occurring k + 1 times each will be increased by one,

and the number of words occurring k times each decreased by one. For equilibrium, the words that had previously occurred k times must be created as rapidly as words that had previously occurred k - 1 times. In this way, the k bin will be replenished as rapidly as it is depleted.
At some point I began to visualize a cascade, with successive pools of water each maintained at a constant level by flow in from the pool above and flow out to the next pool below. Working back from our answerthe distribution that we know describes the phenomenait is not hard to show that the equilibrium condition requires that the probability of creating a word that has already occurred k times must be proportional to k.
We are ready for the final step: to interpret the probability assumption. For word distributions, it can mean that the chance of a word’s being chosen as the next word in a text is proportional, because of association, to how often it has been used already, and also proportional, because of long-term associations stored in memory, to how often it is used in the language. In the case of city sizes, it can mean that birth and death rates are approximately independent of city size, while cities will be visible and attractive to migrants in proportion to their current sizes (Simon 1955b).
I won’t defend these interpretations here. My purpose is to understand the process that reached them. If my account, through the filter of thirty to fifty years of forgetting, has any relation to reality, then we see a process for arriving at the initial formula that looks very BACON-like, followed by working-backward search processes driven by the evocation of prestored mathematical and real-world knowledgeBACON as the front end to an expert system.
Again, my hands are waving wildly. You will not have failed to notice that I have not accounted at all for the cascade metaphor, yet at some time it was evoked and helped me to formulate the steady-state relations. So there is still work to be done on the theory of discovery; still theses to be written and papers published. But I see in this little history, or imagined history, no magic, no mystery. Each step appears to proceed, if not inexorably at least plausibly, from the preceding one.
If the data cried out so loudly for explanation, and if the discovery process proceeded so plausibly, why did not others discover this law and its explanation? Indeed they did. The first was G. Udny Yule, the English statistician, who in 1924 constructed a model very similar to the one I have just described to explain the data on the distribution of species among genera. (I could have been led to this paper by a footnote in Lotka, but I wasn’t.) A second was the English economist D. G. Champernowne, who published “A Model of Income Distribution” in 1953, describing a quite similar process. A third was B. Mandelbrot, who, in 1953, published "An Informational Theory of

the Statistical Structure of Language." I learned about all of these partial anticipations when I searched the literature and inquired among my friends prior to publishing, in 1955, my own paper on the topic (Simon 1955b).
That still isn’t quite the end of the story, for again, the solution of one scientific problem created a host of new problems. In the book I co-authored with Yuji Ijiri, Skew Distributions and the Sizes of Business Firms (Ijiri and Simon 1977), you can find a series of essays applying generalized versions of the same mechanism to understanding the size distributions of business firms and the economic implications of these distributions.
Representations
Mention of the cascade metaphor that I used in finding the law underlying skew distributions raises the question of representations. What kinds of representations do scientists use in thinking about their research problems, and where do these representations come from? One hallowed form of the question is whether scientists (and others) think in words, or whether thoughts take some quite different shapewhether they employ “mental pictures,” for example.
The French mathematician Jacques Hadamard, in his delightful book The Psychology of Invention in the Mathematical Field (1945), comes down heavily on the side of images and against words. Among the many distinguished mathematicians and scientists testifying for him is Albert Einstein, who, in a letter to Hadamard (1945, pp. 142 43), stated that “the words or the language, as they are written or spoken, do not seem to play any role in my mechanism of thought. The psychical entities which seem to serve as elements in thought are certain signs and more or less clear images which can be ‘voluntarily’ reproduced and combined.”
What is good enough for Hadamard and Einstein is good enough for me. I, too, have difficulty in finding any presence of words when I am thinking about difficult matters, especially mathematical ones. Even as I sit here at the keyboard, composing this chapter, I cannot really detect the words in my thoughts (or much of anything else, for that matter) until they come out the ends of my fingers. But perhaps I am not thinking, but just recording previously composed ideas that reside somewhere in my subconscious mind.
Even if we do think in images, neither Hadamard nor Einstein have had much success in describing just what these images are or how they are represented in a biological structure like the brain. Nor have I. I believe, however, that Jill Larkin and I have recently made substantial progress in explaining these matters (see "Why a Diagram Is (Sometimes) Worth 10,000

Words," [Larkin and Simon 1987]). The basic ideas, which I will not elaborate upon here, are that (1) in the course of transforming verbal propositions into images, many things are made explicit that were previously implicit and hidden; and (2) (learned) inference operators facilitate making additional inferences from the images in computationally efficient ways.
We also show, as a by-product of our analysis, that diagrams are representable as list structures, hence are programmable in standard list-processing languages, hence are readily representable in systems of neuron-like structures. Since the surface structures and the semantics of natural languages can also be represented as list structures, it follows that propositions and pictures (or at least diagrams) can use common representational machinery.
Now just as there has long been a debate over whether we use words or images in our thoughts, so has there been a debate (perhaps the same debate) over whether our internal representations of problems look like collections of propositions or like models of the problem situations. Each of these views has been held by an important segment of the cognitive science community, and the two segments do not often communicate with each other, except sometimes to quarrel.
One segment, under the banner “Let language lead the way,” takes verbal reasoning as its metaphor for the problem-solving process, and thinks of reasoning as some kind of theorem-proving structure. The second segment of the cognitive science community uses heuristic search through a problem space (a mental model of the task domain) as its metaphor for problem solving. Human Problem Solving (Newell and Simon 1972) adheres strictly to this viewpoint.
Let me return to my main topic of providing anecdotal evidence about the problem-solving processes used in scientific discovery with an example that I will present rather sketchily, to avoid technical detail.
Economists frequently use what they call “partial equilibrium analysis,” to avoid talking about everything at once by making a host of ceteris paribus assumptions. They examine the impact of a disturbance upon a small sector of the economy while assuming no interaction with the rest of the economy.
If challenged on the legitimacy of this procedure, economists may defend themselves by saying that, of course, interactions are not completely absent but they are small, hence unimportant. We hear that argument not only in economics, but throughout all of science. But is it a satisfactory argument? Small effects, persisting over a long period of time, may accumulate into large effects.
Thoughts of these kinds (represented as words or as images?) went

through my mind while I read, in the early 1950s, a paper by Richard Goodwin, “Dynamical Coupling with Especial References to Markets Having Production Lags,” published in Econometrica in 1947.
Without claiming any clear recollection of the precise steps I took to formulate and solve the problem that his paper evoked, I recall conceiving of a large dynamic system divided into sectors, with strong interactions among the components in each sector, and weak interactions among sectors. I remember also that I worked very hard for several months to see how such a system would behave, and that I worked without paper and pencil while taking long walks.
I held a vague mental image of the matrix of coefficients of the dynamic systemhardly surprising, since this is the way dynamic systems are normally represented in mathematics books. At some point, I saw that the rows and columns of the matrix could be arranged in a number of diagonal blocks with large coefficients in them, and only small coefficients outside the diagonal blocks. The matrix was “nearly block diagonal.” The number of blocks and their sizes were not seen in detail. If forced to give numbers, I might say that there could have been three blocks, each three rows by three columns in sizebut the supposed recollection is surely a fabrication.
At some later point, I acquired a metaphor. I visualized a building divided into rooms, each room divided, in turn, into cubicles. (You can find a diagram of my metaphor on page 212 of The Sciences of the Artificial [1981], 2nd ed.) We start out with an extreme disequilibrium of temperature, each cubic foot of each cubicle being at a different temperature from its neighbors.
Several things now seemed obvious. Throughout each cubicle, a constant temperature would be established very rapidly by the exchange of heat between adjoining spaces. At some later time, each room would attain a constant temperature by heat diffusion through the walls of the cubicles. Still later, the entire building would reach a constant temperature by exchange of heat between the thicker walls of the rooms.
Moreover, because of the differences in the durations involved, each of these processes of equilibration can be studied independently of the others. In studying the equilibration of each cubicle, we can ignore the other cubicles. In studying the equilibration of rooms, we can represent each cubicle by its average temperature, and ignore the other rooms. In studying the equilibration of the building, we can represent each room by its average temperature. As a result, the mathematics of the problem can be drastically simplified.
There still were some difficult mathematical steps from this picture to rigorous proofs of the (approximate) validity of the simplification, but the result to be attained was clear. The reasoning I have described was carried

out mainly in the summer of 1956, and incorporated, together with the mathematics, in a paper written with Albert Ando later that year, but not published until 1961 (Simon and Ando 1961).
I can throw no further light on the source of the heat exchange metaphor, or on how, if at all, I drew inferences from the image of the nearly-block-diagonal matrix. Block-diagonal matrices were not unfamiliar to me, for they had played an important role in my work on causal ordering in 1952 and 1953 (Simon 1952, 1953). The mathematics, which was fairly standard, would have been evoked, I think, in the mind of any mathematician who had put the problem in the form we did.
Our theorems and methods (which may be used to invert matrices that are nearly block diagonal) have attracted the attention of numerical analysts, and of natural scientists who are concerned with hierarchically organized systems. The aggregation method we introduced has also now been recognized to be closely related to the so-called renormalization procedures that play an important role in several parts of physics, and that were invented quite independently of ours.
Even with this sketchy account, the discovery process appears quite unremarkable. The problem was found in the literature (Goodwin’s paper), and was represented in a standard way by matrices having a certain special structure. The metaphor, by showing how such a system would behave, made clear the nature of the theorems to be proved. Although nothing is revealed about the source of the metaphor, it is not at all esoteric. The proofs, while intricate, would not pose any great difficulty for a professional mathematiciana case of normal problem solving, we would have to conclude.
Finding an Explanatory Model
The last two sections provided two examples of the process of finding an explanatory modela model for the rank-frequency relation and a model of nearly decomposable dynamic systems. How could one discover an explanatory model of human problem solving? One way might be by observing some problem-solving behavior closely and inducing the model directly from those observations.
There is a good deal of merit in that answer, and something like that happened when the General Problem Solver was invented. But even in that case, the empirical observations were not the sole source of information that guided the discovery. The inventors also had some notions of the shape of the thing they were looking for.

Explanatory theories take a variety of forms. For example, the behavior of gases is commonly explained by supposing that they consist of a cloud of energetic particles, interacting with one another in accordance with the laws of mechanics. Magnetic attraction between two bodies is explained by a field of magnetic force in the space between them.
One common form of explanation, in both natural and social science, employs systems of differential equations, or difference equations. At any given time, the system is supposed to be in a specified “state,” and the differential equations then determine to what state it will move a “moment” later. Thus, in mechanics, the state is defined by positions and velocities, and the differential equations show how forces produce accelerations that bring about a continuing change in state through time.
Building an explanatory model involves a choice among these or other representations of the phenomena. Will it be a particle model or a continuum model? Will it represent static equilibrium, a steady state, or dynamic change? The representation has to be chosen prior to, or simultaneously with, the induction of the model from the data.
When Allen Newell, Cliff Shaw, and I began to construct a theory of problem solving, around 1955, we were already committed to a representation. In fact, it was our recognition that such a representation had become available with the invention of the digital computer that motivated us to undertake the study of human thinking. We observed that the program of a computer is formally equivalent to a set of difference equations. At each operation cycle, the program determines the new state of the machine as a function of its previous state (the contents of all its memories) together with any new input it has received. Moreover, these difference equations were not limited to manipulating numbers but could process symbols of any kind.
The explanatory task, then, was to describe the processes of problem solving in the form of a computer program. The data we could muster on the behavior of human problem solvers had to be examined for clues to the nature of that program. This requirement provided strong guidelines both for the kinds of data that would be valuable (preferably data that followed the course of problem solution as closely and minutely as possible) and for the best ways of examining the data (searching out the succession of “actions” the problem solver executed and the cues that motivated each action).
Of course, there was more to the representation than simply that it be a computer program. It had to contain symbol structures that could represent the structures in human memory, known to be, in some sense, associative. There was a continuing two-way interaction between the gradual construction of the representation and the construction of the theory that used it.
Sometimes programming convenience (or necessity) dictated choices;

sometimes psychological requirements did. Some aspects of the representation that were initially conceived mainly to meet programming needs (for example, the list-processing languages and data structures in the form of lists and description lists) were later seen to have psychological meaning as networks of associations.
Once some experience had been gained with information-processing models in the form of computer programs, they became a readily available tool for building theories of other aspects of human thinking, as has been detailed in previous parts of this book. No alternative representations were even considered.
In the past few years, with a whole new menu of variants availableproduction systems, models of memory with spreading activation, connectionist models, SOAR, the PROLOG languagechoices of representation have again become an important and difficult part of the model-building process.
Designing Good Experiments
Experiments are supposed to test hypotheses or, better yet, to choose among contending hypotheses (“critical” experiments). That an experiment meet one or both of these aims is neither necessary nor sufficient for its being a good experiment. It is not sufficient because testing weak-tea hypotheses of the form, “variable X affects variable Y,” or its negation is not usually very interesting, and does not often contribute much to our understanding of the world. Testing stronger quantitative hypotheses (for example, the periods of the planets are as the 3/2 power of their distances from the sun) is much more interesting, and very interesting indeed if the hypotheses are closely connected with broad explanatory theories (for example, with the inverse square law of gravitation).
But when we test these stronger quantitative models, we must remember to throw away the whole standard apparatus of statistical significance tests, which is no longer applicable.* We must also remember that models are multicomponent creatures, and when our data do not fit a model, we are faced with the difficult diagnostic task of determining what to changeor whether to discard the entire model.
So much for sufficiency; what about necessity? Is model testing the only

• I cannot pause here to defend this dictum. It will sound like heresy to psychologists but is nearly unanimously accepted by mathematical statisticians. My reasons, and pointers to the literature, can be found in Gregg and Simon (1967).

reason for experimenting? Surely not. One good reason for running an experimentor for spending one’s time just observing phenomena closelyis that you may be surprised. The best things that come out of experiments are things that we didn’t expectespecially those that we would never have imagined, in advance, as possibilities. Of such stuff are many Nobel Prizes made.
Lest I be accused of planning experiments by casting dice, let me suggest that there are heuristics for planning both kinds of experiments, experiments to test models and experiments to generate surprise. (Of course, an experiment designed to test a model may also produce a surprise.) I offer some examples, beginning with a model-testing experiment.
A few years ago, I began to study the Chinese language. I did it just for fun and because I planned to visit China but, to put a more solemn face on things, I called it ''exposing myself to new phenomena." That allowed me to do some of it on company time with a good conscience. Once in China, the Chinese psychologists I worked with and I decided to replicate with Chinese language materials some standard short-term memory experiments. The motive was to test a model. Does Chinese have a magical number (Miller 1956)? And is it the number seven? The answer to both questions was yesno great surprise.
Meanwhile, I had learned a striking fact about the Chinese language (no surprise to my Chinese colleagues, but a surprise to me). A Chinese college graduate can recognize about 7,000 Chinese characters (hanzi). Each character is pronounced with a single syllable. But in the Chinese language there are only about 1,200 distinct, pronounceable syllables (even taking account of tone distinctions). Hence, on average, there are about six homophones for each character.
Somehow (intuition or recognition at work), I remembered that short-term memory (STM) was generally thought to be acoustical in modality, but only because of Conrad’s rather indirect evidence that errors in recall generally involved similarity in sound rather than similarity in appearance. In Chinese, we could put the acoustical hypothesis to direct test. After establishing that the STM span is about six or seven unrelated and nonhomophonic visually displayed characters, we presented the same subjects with strings of visually distinct homophonic characters. The result was dramatic: The STM span dropped to about two or three, confirming Conrad’s result (Zhang and Simon 1985; Yu, Zhang, et al. 1985).
A similar disposition to test models underlies the experiments that Bill Chase and I did on memory for chess positions, building on the earlier work of de Groot and others (Chase and Simon 1973a, 1973b). Could the difference in chess memory between experts and novices be accounted for by

differences in their vocabularies of “chunked” chess patterns? Our experiments demonstrated differences, but not of the right magnitudean answer that, if slightly disappointing, was much sharper than if we had simply asked whether experts’ chunks were larger than novices’.
The experiments on chess memory, like those with Chinese characters, were designed by asking what quantitative predictions a current model made and what measurements would test these predictions. The problem-solving search took place in a task domain, and was facilitated by looking for “surprising” or “interesting” features of the domain. In the Chinese language case, the surprising feature was found first, and the model to which it was relevant was found second. In the chess case, the order was reversed.
These experiments all have an experimental and a control condition, just as any well-designed experiment is supposed to have. In the Chinese language experiments, we compared homophonic with nonhomophonic strings of characters. In the chess experiments, we compared the performance of experts with the performance of novices, and chess positions from well-played games with random positions. The expert/novice dichotomy has also served me in good stead in some more recent experiments in problem- solving in physics (D. P. Simon and Simon 1978; Larkin et al. 1980). An incidental benefit of using this paradigm is that clear-cut experimental and control conditions seem to soothe the savage breast of referee and editor.
Problem Isomorphs
One other experimental manipulation has provided us with almost unlimited mileage: the idea of problem isomorphs. I think I invented the idea of problem isomorphs about 1969, or a little earlier; I do not have any evidence of earlier mention by myself or anyone else. I have a conjecture about its antecedents. (It is a reconstruction, not a recollection, although my Dutch colleague, John Michon, without prompting, corroborated it.)
Saul Amarel was one of the first researchers in artificial intelligence to point out that changing the representation of a problem could sometimes greatly facilitate its solution. Amarel, Newell, and I participated in a semester-long seminar at CMU in 1966 on the topic of problem representation. Now it is only a small step (at least by hindsight) from the idea that a subject can solve a problem easily by finding the right representation to the idea that an experimenter can make a problem harder or easier for a subject by presenting it in one guise or another.
So much for the antecedents. Soon, problem isomorphsproblems with identical task domains and legal- move operators, but described by different

sets of wordswere a topic of discussion in the Understand Seminar (alias the Cognitive Science Seminar), which has run weekly in the Psychology Department at Carnegie Mellon University for twenty years. The first example was number scrabble, an isomorph of tic-tac-toe; and John Michon then added another member to this set. John R. Hayes rapidly became the most prolific and ingenious designer of problem isomorphs, providing us with somewhere between a dozen and two dozen isomorphs of the Tower of Hanoi puzzle, most of which have been used in one or more experiments (Hayes and Simon 1974, 1977; Simon and Hayes 1976).
We have used isomorphs to discover what characteristics of a problem, other than the size of the task domain, account for its difficulty. Early work in problem solving, our own included, had focused on the combinatorial explosion of search as the main source of problem difficulty. Yet we had found that the Tower of Hanoi, with a relatively small and easily exhaustible domain, and the Missionaries and Cannibals puzzle, another much-studied laboratory task, with a tiny one, could occupy human adults for fifteen minutes or a half hour before they found a solution.
The idea that only the size of the task domain could affect problem difficulty sometimes died hard. One referee for a funding agency gave our project proposal low marks, assuming that our experiments could have only negative results, as all isomorphs must be of the same difficulty. (At the time we were told of this objection, we had already demonstrated experimentally differences in the ratio of 16 to 1.)
Experimenting without an Independent Variable
The experiments described up to this point all compare performance under two or more different conditions, by manipulating an independent variable. When I examine my other experimental research, I find to my embarrassment that this fundamental condition for sound experimentation is seldom met. What have I been up to? What can I possibly have learned from ill-designed experiments? The answer (it surprised me) is that you can test theoretical models without contrasting an experimental with a control condition. And apart from testing models, you can often make surprising observations that give you ideas for new or improved models.
Let me start with an example of the latter kind that I have already mentioned briefly. Many summers ago Jeffrey Paige and I took thinking-aloud protocols from high school students solving algebra word problems, in order

to discover what processes they used and to compare their behavior with Bobrow’s STUDENT program, which solved such problems.
Jeff conceived of a fine idea. We constructed some “impossible” problemsproblems that could not be given a real physical interpretation because their solutions involved boards of negative length or nickels that were worth more than dimes. We then asked our subjects to set up the equations corresponding to the problem statements, but not to solve them.
The outcome was wholly unanticipated. Our subjects fell into two groups, rather consistently over a set of three problems. Some set up the equations that corresponded literally with the verbal statements of the problems. Some translated the problems inaccurately, always ending up with equations that described a realizable physical situation. (A few said, "Isn’t there a contradiction?"meaning, “I draw inferences from the problem statements that conflict with my knowledge of the real world.”)
Because we were trying to get as dense a set of data as we could, we had asked the subjects both to think aloud and to draw diagrams of the problem situations. The diagrams drawn by subjects in the first group were generally incomplete and unintegrated, and did not reveal the contradiction. The diagrams drawn by subjects in the second group misrepresented the situations in just the way their equations didso as to make them physically realizable. The direction of the casual arrow is not clear, but one can use these results to conjecture that subjects in the second group used imagery to represent the problem situations before translating into the language of algebra. Subjects in the first group translated directly to equations using only syntax to guide them.
With this kind of information in hand, one can begin to construct models for these sorts of behavior, and to make additional predictions. The ISAAC program, written by Gordon Novak to solve physics problems presented in natural language, uses an internal diagram of the problem situation to mediate between the verbal stimulus and the equations it finally constructs (Novak 1976). The UNDERSTAND program that John R. Hayes and I constructed, around 1972, to show how verbal problem instructions could be converted into inputs for a GPS-like problem solver, borrowed this same insight from the algebra experiments (Hayes and Simon 1974). All of this work was antecedent to the current investigations, mentioned earlier, of representation and imagery.
But the most massive set of examples of the experimental strategy of “just looking” is to be found in Human Problem Solving. Density of data was the name of the game, and protocol analysis the way of playing it. Both Al Newell and I agree that the core of GPS was extracted directly from a particular protocol that we can identify. We also agree in what week in the

summer of 1957 it was done. On the details, the evidence is not wholly concordant, but the main lesson is clear: The GPS theory was extracted by direct induction from the thinking-aloud protocol of a laboratory subject, without benefit of an experimental and a control condition.
What, in addition to luck, entered into the result? First, we already knew that we wished to represent our model as a computer program in a list-processing language. Second, a data-gathering method was used that obtained the densest record of the subject’s behavior we knew how to get. Third, some care had been taken in selecting the task. Application of these criteria to the selection of problem-solving tasks accounts for a substantial fraction of the knowledge that has been collected about problem-solving processes during the past thirty years, and a substantial part of the theoretical efforts that have succeeded in accounting for behavior in many kinds of tasks.
Do these experiments really lack independent variables? Can’t we consider the task domain or the subject to be just that? Or course we can, but why should we? The principal knowledge gained from these experiments did not come out of comparing between tasks or subjects. It came out of painstakingly analyzing individual protocols and inducing from them the processes that problem solvers employed.
Once this had been done, we could test the generality of our results by comparing over tasks and over subjects. But detailed longitudinal analysis of the behavior of single subjects was the foundation stone for the theories we have built.
If the methodology troubles us, it may be comforting to recall that detailed longitudinal analysis of the behavior of a single solar system was the foundation stone for Kepler’s laws, and ultimately for Newton’s. Perhaps it is not our methodology that needs revising so much as the standard textbooks on methodology, which perversely warn us against running an experiment until precise hypotheses have been formulated and experimental and control conditions defined. Perhaps we need to add to the textbooks a chapter, or several chapters, describing how basic scientific discoveries can be made by observing the world intently, in the laboratory or outside it, with controls or without them, heavy with hypotheses or innocent of them.
The Scientist As Satisficer
My economist friends have long since given up on me, consigning me to psychology or some other distant wasteland. If I cannot accept the true faith of expected utility maximization, it is not the fault of my excellent education in economics.

Alas, it did not take. My traumatic exposure in 1935 to the budgeting process in the Milwaukee recreation department had made of me an incorrigible satisficer. I have sketched the theory of scientific discovery to which my study of these problems has led me. It is not a theory of global rationality but one of human limited computation in the face of complexity. It views discovery as problem solving; problem solving as heuristic search through a maze; and heuristic search as the only fit activity for a creature of bounded rationality.
Some scientists believe that theories should be judged by their ability to make correct predictions. I have provided here some tests of the predictive power of this problem-solving theory of discovery. The anecdotes from my own scientific life are instances where it gives a pretty good account of the processes visible in my research.
The problem-solving theory describes me, like KEKADA, formulating a new problem in response to surprise at encountering an unexpected phenomenon. It traces my BACON-like progress toward discerning a lawful regularity in data, and the evocation of knowledge, in expert-system style, to explain the regularity. It accounts for my use of diagrams to gain a grasp of complex phenomena in a dynamic system. It illuminates how the availability of representations and the invention of new ones has influenced my efforts to construct explanations. It characterizes a number of my strategies for designing experiments, and perhaps even explains why I am frequently unconcerned about such things as experimental controls or even independent variables.
Of course, I am exercising poetic license in talking of predictions. A comprehensive Simple Simon has not been programmed; only pieces of him exist. It would be more defensible to talk of explanatory accounts than of predictions. But you will not be misled by the metaphor, which is as useful as one can expect a metaphor to be.
The information-processing theory of discovery that I have been describing has one other virtue. It is not only a descriptive theory but a normative one as well. Not only does it predict (explain) my behavior successfully, but, unbeknownst to me, it has provided me for fifty-three years with a reliable set of heuristics for conducting research. Quite unwittingly, I have been following the instructions of BACON, of STAHL, of GLAUBER, of DALTON, and of KEKADA. I couldn’t have had better guidance.
Even combined together, these heuristics fall far short of a master plan guiding my research career. In any given year, I have seldom known what next year’s experiments or next year’s problems would be. But the heuristics introduced a slight bias into my decision making. Each time I came to a point of choice, they nudged me along one path rather than anothera

little more reliably than if I had tossed a coin. In reviewing my research and my life, that is as much plan as I can find.
One heuristic that has been of first importance to my work is missing, however, from the programs I have described in this chapter: To make interesting scientific discoveries, you should acquire as many good friends as possible, who are as energetic, intelligent, and knowledgeable as they can be. Form partnerships with them whenever you can. Then sit back and relax. You will find that all the programs you need are stored in your friends, and will execute productively and creatively as long as you don’t interfere too much. The work I have done with my more than eighty collaborators will testify to the power of that heuristic.