Reviewed by
Robert
Andersen
Department of Sociology,
University of Oxford.
John Casti's Would-be Worlds is a non-technical introduction to computer simulation modelling. The book essentially consists of short summaries of many projects that have used computer simulation. Although these examples deal with a wide range of topics - from biology to the stock market - there is a significant amount of social science content.
The main argument of the book is that computer simulation modelling represents a major scientific advance because it allows us to study the mechanisms of complex systems: systems that arise from the interaction of their parts. Casti argues that computers can be used to create "artificial worlds". These artificial worlds can then function as laboratories that allow simulated experiments about the real world. Casti's position is summarised in the following quotation:
"For the first time in history we are in a position to do bona fide laboratory experiments on these kinds of complex systems ... now, thanks to the availability of affordable, high-quality computing capabilities, we can actually construct silicon surrogates for these complex, real-world processes. We can use these surrogates as laboratories for carrying out the experiments needed to be able to construct viable theories of complex physical, social, biological, and behaviorial processes. In many ways this leaves us in the same position that physicists were in at the time of Galileo. We now have an essential tool that can be used to create theories of complex systems, theories that will ultimately compare favorably with the theories of mechanical processes that Newton and his successors developed to describe particle systems." (p. 35)
Although Casti acknowledges that a "decent" theory of complex systems is "not even close", he argues that computer simulation will eventually lead the way. The book's five chapters set out to show this to the reader.
The first chapter gives a brief description of computer simulation, model types and model assessment. Although the description is interesting and easy to understand, it does little to help the thesis that computer simulation provides answers to otherwise unanswerable questions. For example, the book starts by discussing a computer game that simulates the NFL tournament (professional American football). Individual agents in the computer game are assigned scores based on the individual statistics of real life players. As in the real game, the team is theoretically as good as the sum of its players' statistics and their interaction. While the example is intriguing, it would be foolish to use the computer game to determine probable outcomes of real football matches because there are many intangibles that it does not include (e.g. coaching strategies, team morale etc). Moreover, even the statistics used in the program are surely not completely reliable, especially for inter-conference or inter-divisional games, since each team plays mostly within its own division and has a unique schedule. Apparently unaware of these limitations, Casti uses the game to assess the outcome of the 1996 Super Bowl between San Diego and San Francisco by performing 100 simulations of a game between the two teams. Very few would have expected the 1996 Super Bowl to be close, something acknowledged by Casti himself, and the betting odds favoured San Francisco by 19 points. The simulation resulted in San Francisco winning only 54% of the matches with a 95% confidence interval of 6.67 (3.93 for the margin of victory, far off the mark of the actual 23 point victory for San Francisco). Rather than acknowledge that the simulation worked poorly, Casti states, "As the final score was 49 to 26 - a 23-point margin - we see what an anomaly, statistically speaking, this game was." (p. 217) It seems here that Casti knows more about the "would-be world" of American football than its "real world". Although not his intent, this exemplifies how results from computer simulations can be misleading when not enough consideration is given to the mechanisms that operate in the real world. Unfortunately, Casti did not seize the opportunity to discuss this.
Despite this example Casti is clearly aware of the problems associated with model assessment. He states, "As with all computer exercises, it's garbage-in, garbage-out, and the faith we place in the model's answers is inversely proportional to the amount of garbage that goes in." (p. 31) He also argues, "In short, the standard of judgment as to whether the model is good or bad is grounded in how well the model answers our questions about the real world of people, places, and things." (p. 46) Unfortunately, however, later in the same chapter he states, "Finally, the realism of a simulation is measured by the realism of the process, as opposed to the realism of the data." (p. 75) It would be unclear to any novice what this means. A little more elaboration would have helped.
Chapter two discusses simulation models that have been used to study the stock market, geopolitics and biological evolution. I agree with Casti that financial markets are good examples of complex systems that are difficult to understand using conventional methods and his discussion of Arthur and Holland's surrogate market is interesting. Since it is impossible to perform real-world experiments on the stock market, computer simulations with realistic assumptions may provide some insight. The chapter is less convincing in its discussion of Richard Dawkins' simulation of how genetic mutation and natural selection interact to make new organisms. Unfortunately, not enough information is provided to understand exactly how the process of developing a computer organism can be used to understand the development of real, carbon-based organisms. Casti's states that what "... immediately catches the eye about this biomorph is its uncanny structural resemblance to a real-life organism called radiolaria, which includes things like the amoeba." (p. 41) This is ridiculous. A completely different biomorph could and would evolve if different criteria were entered into the model.
Chapter three, which discusses the "science of surprise" is particularly good. Here Casti lists five surprise generating mechanisms associated with computer simulation: paradoxes, instability, uncomputability, connectivity and emergence. Paradoxes result from false assumptions about a system leading to inconsistencies between what we expect the model to show and what it actually tells us. Instability refers to the fact that even apparently stable systems can be sensitive to small disruptions. Uncomputability occurs when the processes under investigation are not rule-based. Quite rightly, Casti acknowledges here that there is "... no a priori reason to believe that any of the processes of nature and humans are necessarily rule based." (p. 89) Connectivity refers to findings of unexpected interconnectedness between elements of the system. Emergence implies interactions among elements that are unexpected when the individual elements are considered individually. Also of interest to social scientists in chapter three are the discussions of Arrow's Impossibility Theorem and Arthur's assessment of rational expectations using computer simulation.
The main concern of Chapter four is how learning mechanisms can be incorporated into computer simulation models. This chapter provides the most convincing practical application of computer simulation, called TRANSIMS. TRANSIMS is a program used by traffic planners in Albuquerque, New Mexico to assess the impact of new road construction on traffic patterns. The chapter's discussion of Sugarscape, which models the evolution of cultural phenomena, is also interesting but less believable. Casti is once again too evangelistic when discussing Tierra, a program that emulates the development of a living organism. He states, "January 4, 1990, a day to be remembered: the day when the first noncarbon-based life form came bubbling up out of the computer machine of Tom Ray, a naturalist from the University of Delaware." (p. 162)
Chapter five places the philosophy of computer simulation in the context of scientific knowledge. Casti argues that despite examining surrogate worlds instead of the real world, simulation is a scientific enterprise because information gathering is based on a set of rules that are reliable, objective, explicit and public. The chapter argues that when considering the limits of scientific knowledge we must consider three worlds: the physical, the mathematical and the computational. Casti states, "It is the relationship among these very different universe that must be kept uppermost in mind if there is to be any hope of creating a viable theory of the limits of scientific knowledge." (p. 202)
The book was clearly written as an introduction to computer simulation for the beginner. Still, it does not provide enough detail for even the beginner to understand how these models are constructed and evaluated. Although there is a good bibliography for each chapter, pointing the reader to related studies, the book would have been better if it had fewer examples, giving them much more elaboration. As a result, the main argument that computer simulation is changing the frontiers of science is not very convincing. I am not denying the potential impact of computer simulation. In the light of the evidence he provides, however, Casti has definitely overstated its importance. Despite my criticisms, Would-be Worlds is interesting and does make one think of the vast number of possibilities for which computer simulation could potentially play an important role.
Return to Contents
of this issue
© Copyright Journal of Artificial Societies and Social Simulation, 1999