Chapter 37 Mind's Web Lab

Mind's Web Lab At the same time, Arthur also developed a keen interest in computer experiments during the early days of the Santa Fe economics program. "We use computer programs to do mathematical analysis and theorem proving as if we were doing conventional economics, but because we study rates of increasing returns, learning, and such vaguely defined concepts as adaptation and induction, the problems are often too complex to be Solve it with mathematics. So we have to use the computer to see how things develop. The computer is like a network laboratory, from which we can observe the results of our thoughts turned into actions."

But Arthur's problem was that, even in Santa Fe, economists got nervous at the thought of using computer models. "I guess we're going to have to introduce computer simulations in economics, but I guess I'm too old to change that," Arrow said sullenly over lunch one day. "Thank God, my boy, if the age of theorems is going to pass away, so will I," Hann, who is in his sixties, said on another occasion. Arthur had to admit that economists' skepticism and hesitation were justified.In many ways, he felt the same way himself.He said: "Computer simulations have never been successful in the field of economics. In my own early experience, my colleague Geoffrey McNicaw and I spent a long time observing the role of simulation models in economics." effect, we draw two conclusions, which are now the general consensus. The first is that, in general, only those who cannot think analytically resort to computer simulations. Economics requires deductive computer simulations are just the opposite. The second conclusion is that you can draw whatever conclusions you want from computer simulations as long as you set your own assumptions on computer models. People often start from basic From a political point of view, for example, we need to lower taxes, then make a hypothesis about how beneficial it is to reduce taxes. Jeffrey and I designed a game that can go deep into the model to find out how changing an assumption changes The whole result. Others have done similar things. So computer simulations get a bad reputation in the social sciences, especially in economics, as if they were roguelikes."

In fact, even after all these years, Arthur found himself wary of the word "simulation."He and his colleagues prefer to refer to the economics procedures as "computer experiments," a term that captures a certain rigor and precision practiced by the physicists at Holland and Santa Fe.He says he was greatly inspired by the way Holland and physicists approached computer simulations. "I think it's brilliant. In the hands of extremely careful people, where all the assumptions are carefully considered, the whole algorithm is well-defined, and the simulations are as repeatable and rigorous as a laboratory experiment, in this case, I have found that computer simulations are invulnerable. In fact, physicists are telling us that there are three ways to do science: mathematical theory, laboratory experiments, and computer simulations. You have to use all three over and over again. When you find The results of computer simulations seem to be unreasonable, and you go back to understand it theoretically, and then use the theory as a basis to go back to computer simulations or labs to confirm. For many of us, it seems that economics research can also So we're starting to realize that economics used to be too self-limited and only dealt with problems that could be analyzed mathematically. But now we're in the world of induction, where everything is so complex that we can use computers Experimenting to expand the scope of research. I see this as an imperative, a liberation."

Of course, Arthur hoped that the Santa Fe Economics Project would develop computer models that would convince other economists.Or, at least not to disappoint them again.Indeed, by the fall of 1988, Arthur and his economics project team had used several of these computer experiments. Arthur's cooperation with Holland directly gave birth to the idea of ​​the original glass house economy. “When I arrived at Santa Fe in June 1988, I realized that we needed to start with a real problem rather than start with a full artificial economics model. This idea led to the artificial stock market model.”

Arthur explained that of all the old-fashioned problems of economics, the behavior of the stock market is one of the oldest.This is because neoclassical economics finds Wall Street utterly irrational.It has been argued that since all economic actors are perfectly rational, all investors must be perfectly rational.Moreover, since these perfectly rational investors have the same information about long-run expected earnings for all stocks, their estimates of per-share value, that is, expected net earnings after deducting interest rates, should always agree.According to this reasoning, such a completely rational market can never fall into speculative hype and crash, and the stock market will only fluctuate slightly due to the latest information on various stock expectations.In any case, logically, the floor of the New York Stock Exchange must be a very quiet place.

But in fact, the floor of the New York Stock Exchange has almost become an out-of-control place, with all kinds of bubbles and crashes sweeping the trading floor, not to mention the crowd's fear and insecurity, ecstasy and desire to mobilize riots. It's like a pot of porridge.If a Martian subscribed to an interstellar version of the Wall Street Journal, after reading the newspaper, he might think that the stock market is a living thing.“In reporting on the stock market, journalists always describe it in psychological terms: the stock market is jittery, the stock market is depressed, the stock market is full of confidence,” Arthur said. The stock exchange itself is a form of artificial life.Arthur said, so they thought in 1988 that simulating the stock market in the way of Santa Fe seemed to be the only way to explain the stock market. "The idea is to do a small dissection, modeled first with the assumption of a fully rational actor of conventional neoclassical economics, and then replace it with an artificial intelligence agent that can learn and adapt to the environment like a human. The The model has a stock market where actors can buy and sell freely. As they get a grip on the rules of trade, you'll see all sorts of market behavior emerge."

The obvious question is, what kind of emergent behavior will occur?Would these actors be as comfortable trading stocks as conventional economics describes them?Or will their behavior be as realistic as a turbulent stock market?Arthur and Holland had little doubt that the latter would happen.But in fact, even in the Institute, many people are skeptical. Arthur has a particularly fond memory of a meeting in March 1989.Holland was returning to the institute from Ann Arbor, and several others were present at the economics seminar in the chapel.When it comes to the subject of stock market models, both Sargent and Lemon Marimon of the University of Minnesota argue vigorously that the price bid by the adaptive agent will quickly converge toward the "base price" of the stock, that is, it will definitely This is what neoclassical economic theory predicts.The stock market may have occasional ups and downs, but the actors can't really do anything else, they say.The base price attracts them tightly like an infinitely huge gravitational field.

"John (Holland) and I looked at each other and shook our heads. We said, that's not possible. We had a strong intuition that the stock market we'd built had enormous potential to behave in a self-organizing way that would grow more and more complex. New, rich behaviors emerge," Arthur said. They had a heated argument about it, Arthur recalls.Arthur knew, of course, that Sargent had been fascinated by Holland's way of learning since his first economics seminar in September 1987.In fact, Sargent began to study the impact of learning on economic behavior long before this.And Marimon was as interested in computer experiments as Arthur was.But in Arthur's view, Marymount and Sargent did not really seem to regard learning as a new perspective for studying economics.They seem to use learning as a way of consolidating conventional economic theory, as if it were a way of understanding how economic agents grope for the behavioral patterns of neoclassical economics without being fully rational.

In fairness, Arthur must concede, the two men had reason to think so.Sargent had some experimental evidence to support their thesis, in addition to the "rational expectations" theory he worked on.The researchers demonstrated that in a series of computer simulations in which students acted as stock traders, the trading prices of the subjects quickly converged to the stock's underlying price.And, Marymount and Sargent are running their own Santa Fe-esque computer simulation: an old problem known as the Wicksell triangle.The plot goes like this: three different types of actors produce and consume three different types of goods, one of which eventually becomes a medium of exchange: money.When Marimon and Sargent replaced the rational actors in the original model with a system of classifiers, they found that the system every time returned to the conclusion of neoclassical economics (that is, the medium of exchange is the goods with the lowest inventory cost — for example, a metal disk instead of fresh milk).

Nevertheless, Arthur and Holland still did not give up their efforts."The question is, does realistic adaptive behavior actually lead to a rationally desired outcome?" Arthur said. is possible. Fundamentally, the theory of rational expectations says that humans are not stupid. It's like playing tic-tac-toe and learning to predict your opponent's behavior after a few tries behavior, so that both parties can play the game seamlessly. But if it is a situation that will never be repeated, or in a very complex situation, the actor must do a lot of calculations, then your requirements for the actor It's too high. Because you're asking them to understand their own expectations, to grasp the driving forces of the market. To grasp other people's expectations. And other people's expectations of other people's expectations, and so on. In this case, Arthur and Holland believe that the actors will be in an extremely unbalanced situation, and the "universal gravitation" that leads to rational expectations will become very weak, and the motivation and Accidents rule everything.

The argument, which was both friendly and heated, continued for some time, Arthur recalls.Of course, neither side backed down in the end.But Arthur clearly felt this was a challenge for him: If he and Holland believed their stock market simulations would exhibit realistic emergent behavior, then they had to prove it. Unfortunately, the programming of the stock market model has been intermittent. Over lunch one June day in 1988, Arthur and Holland sketched out the first steps of the computer simulation.They were both lecturers at Santa Fe's inaugural summer class on complex systems.That summer, Holland returned to Ann Arbor and wrote a complete classifier system and genetic algorithm in the only computer language Arthur knew, BASIC (which finally freed Holland from using hexadecimal notation Writing programs. He had to teach himself BASIC language. But since then he has only written programs in BASIC language).That fall, during the first few months of the economics program, Holland returned to Santa Fe.Once he returned to Santa Fe, he worked with Arthur to further develop the stock market model.But because writing this as an ecosystem took up a lot of Holland's time, and Arthur was struggling with administrative affairs, the stock market model writing progressed slowly. To make matters worse, Arthur began to realize that although the concept of a classifier system had many strengths, it was cumbersome to use."Initially, people in Santa Fe thought the sorter system was the answer to the stock market and could make you coffee in the morning," he says. "So I used to joke with Holland, 'Hey, John, sort Is this system really capable of producing low-temperature fusion?'” "By early 1989, David Lane and Richard Palmer organized a group devoted to Holland's thought, which met four times a week before lunch. Holland had left Santa Fe by then, but We spent a month studying his book "Induction". When we went deep into the technology of the classifier system, I found that the architectural design of the classifier system must be very careful to ensure that it is practical. It can be used in , and the link between the rules must be carefully designed. At the same time, you may design a 'deep' classifier system, that is, one rule activates another rule, and then activates another A rule, thus causing a long chain reaction of classifier systems, or you may design 'breadth' classifier systems, i.e. stimulus-response systems, which can produce 150 way of responding. My experience is that wide systems learn very well, whereas deep systems do not.” Arthur and Holland's former student, Stefania Forrester, have done much to explore this question.Forrest is now at the University of New Mexico and is a regular at the Santa Fe Institute.She told Arthur that the problem was Holland's bucket queue algorithm.This algorithm can reward all kinds of rule theory.If the bucket queue algorithm can be reversed to reward the previous generations of rules, then by the time the originator of these rules is traced back, the rewards will be running out.So it's no surprise why the learning function of the shallow system is better.Indeed, the refinement and improvement of the bucket queue algorithm has become the most urgent link in the research of the classifier system. Arthur said: "These things made me suspicious of the classifier system. As I became more familiar with the system, the disadvantages of the system became more and more obvious. However, the more I looked at this system, the more I marvel at the thought involved: you can have many conflicting assumptions in your head that can compete with each other, because then you don't have to preprogram some kind of expert into the system, which I really appreciate The idea. I started to think of Holland's systems from a slightly different angle than Holland's. I thought of them as ordinary computer programs, with many moduli and branch points, but each time the program had to decide for itself what to do. which modulus to activate instead of activating the modulus along a fixed sequence. Once I started to think of them as a self-adapting computer program, I felt much smoother. I think this is exactly what Holland where the achievement of virtue lies." Anyway, they finally finished their version of the stock market model, he said.Sargent's many simplifications to the original design helped a lot in this version. In the late spring of 1989, Richard Palmer, a physicist at Duke University, also joined in and strongly supported the introduction of this model with his superb programming skills. At the same time, Palmer, like Holland and Arthur, was fascinated by the model.He said: "This model is about self-organization, which is an area of ​​research that I am deeply interested in. How is the brain organized? What is the nature of self-consciousness? How does life spontaneously arise? I have been circling these in my head. Big question." In addition, he was anxious about another Santa Fe research project that he had already spent a lot of time working on.The project is the "Double Outcry Competition" model, which he developed with John Miller of Carnegie Mellon University and John Rust of the University of Wisconsin.The competition, which was eventually held in early 1990, was conceived at the first economic symposium in September 1987.This model is very similar in principle to the one Axelrod devised 10 years ago.But instead of replaying the "prisoner's dilemma" game, this model includes a variety of strategies traders use to deal with commodity markets like stock trading.Is it best to call the price as soon as the market opens?Do you just keep quiet and wait until the best price comes along before you bid?Because buyers and sellers bid spontaneously in such a market, the system is called "double bidding," and the answer is unknown. The racing game should be a lot of fun, Palmer said, and programming it was certainly a big challenge for him and his colleagues.But the actors in this model are largely static.To him, the racing game simply didn't have the magic of the Arthur and Holland German models.In Arthur and Holland's model, you can see the actors becoming more and more complex, able to develop into their own real economic life. Palmer has devoted all his energy to the design and development of the stock market model since early spring. In May 1989, he and Arthur completed the first version of the stock market model.According to the intention of their design, at the beginning of this model, its actors are completely ignorant, and they are all arbitrary rules, allowing them to learn how to bid by themselves.They found that the actors learned as rapidly as they expected. They observed that every time the system was run, the results were in line with Tom Sargent's predictions as if they had seen a ghost.Arthur said: "In this model, the dividend of a single share is three dollars, and the discount rate is ten percent. In this way, the base price of the stock is thirty dollars. And the stock price really fluctuates around thirty dollars, proving that Conventional economic theory is correct!" Arthur was deeply frustrated and troubled.The only thing to do now seemed to be to call Sargent back from Stanford and congratulate him on the win. "But Richard and I walked into the office one morning and ran the system on my Macintosh. We've been watching it work and we've been talking about how to improve the program. We noticed that every time the price hit thirty-four dollars , the actor buys in. We can visualize this situation, and it seems unusual. We thought the model was wrong. But after another hour of hard thinking, we realized that The model is not wrong, but these actors have discovered the original form of technical analysis. That is, these actors start to believe that if the stock price rises to a certain level, it will continue to be bullish, and then buy. But of course, this understanding It becomes a kind of self-fulfilling prophecy: If there are enough players willing to buy when the stock reaches thirty-four dollars, it will cause the stock price to continue to rise." Moreover, when the stock price fell to twenty-five dollars, the exact opposite happened: the players would sell as much as they could, thus forming a self-fulfilling prophecy about the stock market being bearish.This is exactly why stock market bubbles and crashes occur!Arthur was so excited that even the most cautious Palmer was infected by his enthusiasm.This conclusion has been confirmed repeatedly in newer, more complete versions of the model, Arthur said.But on the morning of May 1989 they realized they had succeeded. "We realized right away that we had glimpsed a glimmer of an emergent feature in this system, a glimmer of life."