Footnotes on things I’ve been reading: Epstein and Axtell, “Growing Artificial Societies”

Epstein and Axtell, Growing Artificial Societies
(1996).

I've become very interested in the "artificial societies" approach to social science. The idea is that we attempt to understand the workings of a society not through "discursive" models (as in qualitative social science) or through game theoretical models with extremely simplified assumptions and homogeneous agents (as in more "quantitative" approaches), but by creating a simplified version of the society in silico. So we use software (or pen and paper; Schelling's pioneering models of segregation were studied without the benefit of a computer) to generate an environment and simplified agents that interact with one another, and then we see what happens. The advantages of this approach is that one can add complexity to a model in a controlled manner and study the resulting dynamics (rather than merely static equilibria); so, for example, one might start with agents who move and eat, then see what happens if you add the ability to trade, and so on (besides, it's fun to play around with such models). The disadvantages, however, come from the very flexibility of the approach, which allows for highly complex models: it is sometimes hard to tell whether or not a particularly interesting result is simply an artifact of the simulation or the consequence of some particular simplifying assumption (though it should be noted that the same is often true of traditional models, where striking results, like the pareto-optimality of free markets, are sometimes basically artifacts of the simplifying assumptions made for the sake of creating "tractable" models).

Epstein and Axtell report in this book the results of a pioneering 1994 simulation: "Sugarland". They start with a very simple environment, the "sugarscape", consisting of a two-dimensional surface (technically, a torus) with an unevenly distributed resource ("sugar") and add some very simple agents who can move around the sugarscape according to simple rules and "eat" the sugar. As they add more complex rules, they basically "generate" a number of features of societies – like migration patterns, cultural transmissions, combat, trade, credit relationships, disease transmission, etc. – and study how some simple rule changes affect these patterns. The book does not describe in detail how to do this (they give some information that might help in replicating the simulation, but no code); instead, one can read it as a kind of advertisement for the generative approach to social science. Epstein and Axtell excitedly argue that many phenomena can be understood as "emergent" effects of the interaction of agents following simple rules, and hence that the best way of understanding them is by looking at which simple rules are able to generate them. In general, I am very sympathetic to this idea; it is unlikely that we will ever fully understand polities and economies through models composed of homogeneous, fully rational agents.

But though Epstein and Axtell's results are often suggestive, it is not always clear that they have properly looked at the ways in which minor changes in the parameters of the simulation might disrupt the patterns they find, or sufficiently thought about how to interpret the results of their simulations. As a result, the book is somewhat disappointing: if you are already convinced of the utility of the generative approach, you may not learn much here, and if you are not, then you may think that the authors have not really addressed the important objections to the generative approach. A good example is in the chapter of the book on combat and cultural transmission. Here the assumptions made about how to model cultural transmission and combat between agents seem rather arbitrary (rather than based, for example, on research about human or animal combat), and though they explore a number of alternative specifications, the results seem only lightly grounded theoretically, more an artifact of the simulation than an illuminating model of actual cultural transmission or combat. (By contrast, the chapter on trade and credit is probably the most solid in the book, since they ground the results in economic theory and systematically explore some of the space of alternative rule specifications and parameter values for their agents. Nevertheless, their results there do not go beyond what most sophisticated economists already knew, and do not add much knowledge about the properties of dynamic adjustment in markets – we get no bubbles or other interesting phenomena, for example, and we would need to introduce something like production functions to make the model better reflect actual economies).

Of course, the book is 16 years old right now, and new advances have been made in agent-based simulation. Besides, as I mentioned above, the book is perhaps best read as an advertisement for the virtues of the "artificial societies" approach rather than as a contribution to the study of actual societies; it opens your appetite rather than satisfies it (I wanted to do some programming as I was reading it, to try to see if I could replicate their results). But I think it is fair to say that the book does not quite succeed at showing that the artificial societies approach is actually worth taking seriously (unless you are already convinced of its merits).

I suspect that in order to make the "generative" approach work better, we need to be clearer about how the simple rules that agents follow in the simulated world relate to the kinds of rationality displayed by real political and economic agents than this book is. I also wonder how we could use this approach to model political and not merely economic phenomena. This is something I would like to do: how would you model the emergence of the state, or the choice of political regime? Do you need to incorporate more rationality into the agents, or a way of representing social conventions? Anyway, all in all, I enjoyed reading the book, and people may find it gives them new ideas, if nothing else.