Book Review
Artificial Life 3:3: 237-242 (1997).

Nicholas Gessler
Department of Anthropology
University of California at Los Angeles.
Computational Evolution and Ecology Group
Center for the Study of the Evolution and the Origin of Life
c/o 11152 Lucerne Avenue
Culver City, CA 90230-4244

Growing Artificial Societies - Social Science from the Bottom Up.
Joshua M. Epstein and Robert Axtell. (MIT Press 1996) $39.95 hardcover; $18.95 paperback.

Immersed in the simulations which Epstein and Axtell have created, my mind turns back a quarter century ago, to 1971, the year Stanislaw Lem wrote his own review of Professor Dobb’s related work entitled Non Serviam:

[Personetics:] At present a "world" for personoid "inhabitants" can be prepared in a matter of a couple of hours. This is the time it takes to feed into the machine one of the full-fledged programs… (p. 168) A specific type of personoid activity serves as a triggering mechanism, setting in motion a production process that will gradually augment and define itself; in other words, the world surrounding these beings takes on an unequivocalness only in accordance with their own behavior... (p. 172) As hundreds of experiments have shown, groups numbering from four to seven personoids are optimal, at least for the development of speech and typical exploratory activity, and also for "culturization." On the other hand, phenomena corresponding to social processes on a larger scale require larger groups. At present it is possible to "accommodate" up to one thousand personoids, roughly speaking, in a computer universum of fair capacity... (p. 77) studies of this type, belong... to… sociodynamics... There have arisen many different philosophies (ontologies and epistemologies), and also "metaphysical experiments" of a type all their own… (p. 186)… "I can enlarge their world or reduce it, speed up its time or slow it down, alter the mode and means of their perception; I can liquidate them, divide them, multiply them, transform the very ontological foundation of their existence. I am thus omnipotent with respect to them…" (p. 195)

Written in the same years that Thomas Schelling was pioneering his own work on Micromotives and Macrobehavior, Dobb’s book, by contrast, is apocryphal, as is Dobb himself. Lem’s playful review of a non-existent book is nonetheless insightful, and he and Schelling deserve credit for their remarkable foresight. Today the U.S. Army Simulation, Training and Instrumentation Command (STRICOM) casually displays the motto "All but war is simulation." Its Distributed Interactive Simulation (DIS) project’s logo integrates "live," "virtual" and "constructive" processes with a "synthetic environment." The entertainment industry is bringing artificial life into our homes with increasingly anthropomorphic agents. BioBattler (Victor Company of Japan) subtitled "Get God’s Hand!" is an apocalyptic vision of space colonists who "create the special DNA of mutant pets and let them kill each other." Creatures (Cyberlife, Warner Interactive, London) by contrast, is a compellingly cute and welcoming world where you create and breed life "known as Norns… (who) by nature, are a friendly, curious, social species, eager to learn and experience life to its fullest." They even evolve their own language. Disbelief is suspended to such an extent that its creator, Stephen Grand, received a dying Norn from one of its guardians requesting that he cure it. He did. It took an entire day. He later quipped, "I don’t know who was the more foolish." The market for simulations is steadily growing. Something is in the air.

Growing Artificial Societies is a milepost publication. It is the first book in the last two decades, aside from conference volumes, to suggest situated multiagent models as research labs to study serious issues in social science. As a friend, Sharon Traweek, recently reminded me, "More men have walked on the Moon than have stepped into artificial societies and artificial cultures." The authors, both from Brookings Institution, present us with their portable society Sugarscape, developed in collaboration with the Santa Fe and the World Resources Institutes as Project 2050. Their book is clearly written, in easily accessible simple language. Laying out assemblies step by step and piece by piece, they invite us to dig in, prepare a foundation, frame homes for a world and people it. They present a wide range of suggestive experiments which expose the normative assumptions of traditional formalisms in social science as unlikely special cases. They target economic theory, but their critique has added implications for fields such as anthropology, where formalisms have been much less influential. In their guidebook to a very foreign land, we’re led in six successive chapters across its ecosystems and peoples. We sample histories of its citizens and their quarrels, sex lives and cultural habits. Finally, we see them trading and getting sick. All this, and they’re not even conscious of what they’re doing. At the tour’s end are conclusions, appendices and useful references. Travel papers in order? A CD-ROM is expected.

Epstein and Axtell introduce us to the argument that crucial social processes are complex, not neatly decomposable into disciplines. They remind us that the social sciences have reduced the individual’s relation to society through the construction of the average person. They challenge the assumption of shared global knowledge and rationality. The solution they propose is one step up from artificial life:

We view artificial societies as laboratories where we attempt to "grow" certain social structures in the computer --- or in silico – the aim being to discover fundamental local or micro mechanisms that are sufficient to generate the macroscopic social structures and collective behaviors of interest. (p. 4)

After acknowledging preceding research, they introduce us to the geography of silicon worlds, the demography of their agents and their environments. The world they offer, instantiated through object oriented programming (OOP) is their Sugarscape, their CompuTerrarium. They prime us to look for history, family trees, tribe formation, combat, trade, credit and disease. They give the take-home messages up front: heterogeneity (not normative aggregate individuals), space (not point kinetics), simple rules to global patterns (not top down causation), evolutionary dynamics (not equilibria) and emergent (not presumed) higher level regularities.

The broad aim of this research is to begin the development of a more unified social science, one that embeds evolutionary processes in a computational environment that simulates demographics, the transmission of culture, conflict, economics, disease, the emergence of groups, and agent coadaptation with an environment, all from the bottom up. Artificial society-type models may change the way we think about explanation in the social sciences. (p. 19)

This bottom-up strategy has deep implications for philosophy, suggesting not only how we come to know the world around us, but the possible limitations to our knowledge and strategies for building increasingly reliable representations of the external world, a contemporary definition of the scientific enterprise. The scheme conceives the world to be built from bottom-up, but as it reaches each new height emergent properties are captured which reach down to lower levels making top-down perception and control possible. Top-down circuits reach below, but only from elevations as high as bottom-up processes have elevated them. In such ways, our own perceptions of the world are changed by previous perceptions. We fashion our societies after they have fashioned us and we try to make the world accord to our conceptions of it. Theirs is the eminently practical approach of modeling. To show how the system works, they build it and try it out, tinker with it, tweak it, and push it to its limits. For them, this process of modifying and testing constitute explanation. They treat modeling as theory and practice, discourse and experiment.

Perhaps one day people will interpret the question, "Can you explain it?" as asking "Can you grow it?" (p. 20)

The Sugarscape’s geography is a 50x50 toroidal grid of 2500 cells. Its first resource, sugar, ranges initially from 0 to 4 units per cell distributed in two heaps in opposite corners. Its first random population of 400 agents is heterogeneous, having metabolisms requiring daily rations of from 1 to 4 food units, and visions penetrating from 1 to 6 cells in orthogonal directions. Agents look around for the greatest nearest food, eat their ration and appropriate the rest. The replenishment of food may be immediate or delayed over four time steps. Visualizations show the population reach its carrying capacity of 224 agents. At birth, agents are given random life spans. Upon their death they are replaced by newborns with random metabolisms and visions. A Lorenz curve tracks wealth and a Gini coefficient measures economic inequality. Social networks are visualized. Agents’ visions are increased to 10 and these new agents are all plopped down in one corner of the world. Confirming expectations that this world should be believable, some agents migrate to the other side. Next we alter the growth of the sugar piles so that they alternate through time. Sure enough, we get seasonal migration and a sort of speciation into migrators and hibernators. Introducing pollution to dissuade agents from remaining in high food concentrations, we note its repercussions, and then again as we let the pollution diffuse. Having exercised this world, its environment, its agents and its visualizations, as well as our powers of interpretation - despite its heavily grained structure - we can now accept this world as plausible and we are prepared to suspend our disbelief in hopes of witnessing some additional human-like behaviors.

A history of Sugarscape is a "very simple caricature… a ‘proto-history.’" Insantiating sex, defining fertility and implementing simple Mendelian cross-over, we view not-too-surprising animations of the selection for agents of higher vision over low and lower metabolism over high in this "theory of evolution brought to life." They corroborate the oscillations in populations where fitness is dynamically determined on the fly. All of this can be seen in artificial life simulations, so we now introduce human nature into the agents. Inheritance of the acquired character of wealth "retards selection." They introduce a graphic overlay of genealogical networks. With the ingredients in place, we’re ready to start cooking with culture. The culture of each agent is represented as a binary tag string. Each agent at each move has the chance to examine one tag, chosen at random, from each of its neighbors, and flip it to coincide with its own. This represents horizontal cultural transmission. Sub-culture membership (in my opinion something of a misnomer) is decided by the majority, of 1s or 0s, that a tag contains. Vertical cultural transmission from parent to child is again Mendelian. Friendship (which I would describe a sub-cultural affinity) is measured by the closeness of two agents’ tag strings taken bit by bit - the Hamming distance. Combat is conceived as the ability to kill an opposing tribesman and reap a reward without immediate retaliation. The reward itself is varied. First agents are allowed to kill and rob. Then they are turned into bounty hunters. All this leads to an admittedly "crude caricature of early social history." It’s easy to find fault with the authors’ various constructions and interpretations. The authors are economists. But such criticism misses the point. These worlds may be constituted in any manner whatsoever. It is hard to resist the gnawing feeling that the body of theory in other social sciences is in grave need of similar algorithmic reanalyzes. It is up to sociologists, anthropologists and other readers to constitute their own worlds as they see fit.

The authors are now ready to tackle a serious problem, the origin and maintenance of trade and prices as seen by economics. Here they evaluate the consequences of the assumptions of neoclassical microeconomic theory, that agents are ageless, consistent and truthful:

We find that the neoclassical agents trading bilaterally are able to approach, over time, a price close to that associated with an optimal allocation. However, when the agents are made progressively less neoclassical --- when they are permitted to sexually reproduce or have culturally varying preferences --- the markets that emerge generally have suboptimal performance for indefinite periods of time. (p. 95)

Credit and interest, again of paramount importance in economics, motivate further experiments:

When agents are allowed to enter into credit relationships with one another --- for purposes of bearing children --- interesting financial networks emerge. Some agents end up as pure lenders, others as pure borrowers, and many turn out to be both lenders and borrowers. Indeed, entire financial hierarchies emerge within the agent society. (p. 95)

Disease is the last chapter, an investigation into treating inter-agent epidemiology as an emergent property of intra-agent immunology under a variety of conditions. The authors conclude by setting out a time-honored minimalist strategy for research: to obtain the greatest leverage from the smallest lever.

The aim is to provide initial microspecifications (initial agents, environments, and rules) that are sufficient to generate the macrostructures of interest. (p. 177)

In answer to the question, "what sort of science" they are doing, they reply by proposing the term "generative." And they set out what is, in effect, their computationist manifesto:

We are proposing a generative program for the social sciences and see the artificial society as its principal scientific instrument. (p. 177)

Periodically, we revisit deeper issues of epistemology and ontology, issues of the limits of knowledge. If there are theoretical limits to understanding computational modeling systems, what does that imply about the limits to more natural modeling media such as cognition or discourse? Do these limitations apply across the board?

A deeper issue for social science is that there may be theoretical limits to what is knowable in such computational systems as artificial societies… In some areas, it may be that simulation really is the best we can do. (p. 178)

Already, it appears that artificial societies will require many of the same analyses as their natural counterparts. There is one marked advantage to the artificial, the omniscient stature of the researcher. Within the artificial worlds’ bounds we can come to know everything. We can, by diverse means, be anywhere or everywhere at once. Faced with such complete knowledge, too much to cram into our heads, we are forced, in part, to return to our initial problem: how to leverage understanding out of this world through simpler models and propositions. The problem is still one of building increasingly leveraged representations of complex worlds.

What are the evolutionary implications of their study? A specter is haunting social science --- it is the specter of computation. It challenges traditional views of causation and the nature of reality. What we know of the world, we learn from models inside our heads. In this sense the world that we know is artificial, because it is not the world itself. Our heads, though conveniently portable, provide too constrained a space for computation, so we’ve expanded our minds by externalizing some of our thinking. We have excelled in the artifice of making tools for over 2,000,000 years. For the last 35,000 we have made external representations, models and art. Society itself is partly a cognitive artifact, it too being artificial. So what is really new? Recently, our cultural evolution has made an unprecedented leap towards leveraging information. We have captured our creator, evolution, and we have embodied it in our machines. Just as natural selection merged with genetics in the modern synthesis, we are now combining evolution with computers in, what might similarly be called the postmodern synthesis. We have breathed qualities of life into our artifacts as computation. The haunting question is, How much of our own thought and decisions are we willing to entrust to artificial societies of autonomous, evolvable and emotional multiple agents?

On the brink of a new millennium, we would do well to reconsider the domain and quality of the of mental, discursive and mathematical models we have customarily utilized in the social sciences. The authors provide us with compelling reasons to do so. Science moves neither smoothly nor evenly. Like technology, both ongoing parts of culture, it shares with its biological base analogs to punctuated equilibria and to mosaic evolution. Social science is still heavily invested in more traditional modes of explanation and although a thoroughly informed anthropological study of the coevolution of ecologies of various models among researchers in differing environments would be extraordinarily useful, glimpses of what it might reveal remain only vague subtexts in the authors’ work. Computer science is now in its fiftieth year, and it has provided us with languages that enable us to speak clearly about complexity and analyze it. This has never happened before in human evolution. Epstein and Axtell are among the first to introduce the state-of-the art in computer science into social theory. "Growing Artificial Societies" is a must-read for anyone interested in building and testing theories of cultural process. I share their hope that others will soon follow, individually creating ensembles of tenable artificial societies from which we can begin to explore the potential risks, costs and benefits of competing social policies, present, future and past.

We have only begun to explore the uses and limits of the artificial society as a scientific tool. (p. 178)

Rather than stand on the shore watching the ships recede below the horizon, Epstein and Axtell have taken the wheel. They have set one of many courses for us to follow in these new voyages of exploration and discovery to yet more foreign and compelling worlds.


1. Lem, S. (1983). Non Serviam. In S. Lem, A Perfect Vacuum. Translated by M. Kandel. New York: Harcourt Brace Jovanovich.

End of citation.

[ Meta-Knowledge Engineering Server ],
[ Personoids' Organization & Society ]