Agent-based models in anthropology: growing societies from simple rules

An agent-based model (ABM) is a computer simulation built from individuals rather than averages. You describe some agents (people, households, herds, villages), the world they live in, and the rules each one follows. Then you let them act and watch what the whole population does. Nothing at the level of the society is written into the code. If a pattern appears, such as a cluster of settlements, a shared custom, or a famine that hits some households and spares others, it has come out of the interactions.

That is why anthropologists use them. The recursions in our explainer on transmission biases describe a population with one number, the share holding a trait. An ABM can keep track of who lives next to whom, whose granary is full and who copied whom, which is the kind of detail ethnographers and archaeologists actually record.

Four parts, every time

However large the model, it has the same parts:

The schedule is the part beginners skip and critics ask about first. Whether households share before or after they eat can change the result.

A model you can run with a pencil

Six hypothetical households, A to F, sit in a ring. A's neighbors are F and B, B's are A and C, and so on round to F, whose neighbors are E and A.

Each season, apply these rules to every household at once:

1. Harvest: add this season's harvest (table below) to the household's store.
2. Share: any household holding 4 or more units gives 1 unit to each neighbor holding 0 or 1. Decide who gives and who receives from the holdings just after the harvest, before any gift moves.
3. Eat: each household eats 2 units. A household holding less than 2 eats what it has, goes short for the season, and ends at 0.

Every household starts with 1 unit. The harvests came from a die (a roll of 1 gives 0 units, 2 gives 1, 3 or 4 give 2, 5 gives 3, 6 gives 4), which averages exactly 2 units, the amount each household eats:

Here is season 1 step by step. After the harvest, holdings are 3, 3, 3, 5, 1, 2. Without sharing, everyone eats 2: A, B and C end with 1, D with 3, F with 0, and E, holding only 1, goes short. With sharing, D holds 5 and its neighbor E holds 1, so D gives E one unit (C, at 3, does not qualify). Holdings become 3, 3, 3, 4, 2, 2, everyone eats, and stores end at 1, 1, 1, 2, 0, 0. Nobody goes short.

Carry on the same way, each season starting from the stores the last one left, and you should get this:

Count the results. Without sharing, households went short six times in four seasons; with sharing, once. But the sharing village ends with 7 units in store against 12. Sharing did not make food. Both villages had the same 54 units (6 to start plus 48 harvested); the sharers ate 47 of them and the others ate 42, because a household that goes short eats less. Sharing moved food out of the stores of B, D and E and into the meals of their neighbors.

Look also at F in season 2. F went short even with sharing, because neither of its neighbors (E with 2, A with 1) had plenty that season. Whether help arrives depends on who your neighbors are and how their season went, an effect that a village-wide average would hide.

One run is an anecdote

The tables above are a single draw of the dice, and a lucky one for the sharers. Run the same rules 10,000 times on a computer with fresh rolls and the averages are less dramatic: about 3.8 hungry household-seasons per four-season run without sharing against about 2.1 with it, and end stores of about 10.8 units against 8.5. The direction holds, but part of the gap in the hand example was chance. Published models report results over many runs for the same reason, and readers should look for the spread, not just the average.

Growing an explanation

Joshua Epstein called this way of working generative social science. His 1999 paper in Complexity put the motto plainly: "If you didn't grow it, you didn't explain its emergence." On this view, to explain a social pattern you show a set of individual rules that is sufficient to produce it from the bottom up, with agents who differ from one another, act on local information and are not coordinated from above.

Epstein was careful about the limits. Growing a pattern makes a set of rules a candidate explanation, and possibly a weak one. Several different rule sets may grow the same pattern, and then you have to go back to the individuals to decide among them. His warning example was Barnsley's fern, a short mathematical recipe that draws a convincing fern but says nothing about how ferns grow.

The sharing ring shows both halves. It grows a pattern (fewer hungry seasons, smaller reserves) from household rules. But storing more, farming scattered plots or borrowing could grow similar numbers, and only ethnographic or archaeological evidence could say which rule a real community followed.

Three models worth knowing

Long House Valley

The best-known archaeological ABM is the Artificial Anasazi model of Long House Valley in northeastern Arizona, built by a team that included the archaeologists Jeffrey Dean and George Gumerman and the modelers Epstein and Robert Axtell. Its environment is reconstructed from tree rings, pollen and the history of the valley floor, which gives year-by-year estimates of how much maize each patch could yield. Its agents are farming households that choose where to farm and where to live, store food, and split to form new households. Axtell and colleagues (2002, PNAS) reported that the model reproduced the main features of the valley's population history between about AD 800 and 1300, including its rises and falls, its settlement patterns and a steep late decline. As Epstein summarized an earlier phase, the environmental rules produced a sharp drop around 1300 but not the outright departure, which suggested that subsistence alone does not fully explain why people left.

Marco Janssen (2009) rebuilt the model and found that its close fit rested largely on two parameters that set how much food the land produced. A plain carrying-capacity calculation, with no agent behavior at all, tracked the historical population almost as well; the households mostly smoothed the environmental signal. That does not make the model worthless. It shows what the agents were and were not adding, which is the question every reader should ask.

Water temples in Bali

J. Stephen Lansing and James Kremer (1993, American Anthropologist) modeled the subaks, the farmers' irrigation associations of Bali, each choosing a cropping schedule. Planting at the same time as your neighbors helps control pests, but if too many subaks along a river plant together, water runs short. In the model each subak, at the end of a year, compares its harvest with its neighbors' and adopts the schedule that did best. Within a few simulated years the subaks fall into coordinated clusters, and those clusters closely resemble the real groups that coordinate their irrigation through networks of water temples. The model suggested that institutions often described in religious terms were also doing ecological work, and that such coordination could arise without a central planner.

Herders and farmers

Lawrence Kuznar and Robert Sedlmeyer (2005) built NOMAD, a flexible model of interaction between mobile pastoralists and settled farmers that can be adapted to different environments and social settings, and applied it to conflict over land and resources in Darfur. Our topic note on that model explains how it works and how to read its results with care.

Seven questions to ask of any model

1. What is it for? A model meant to explain a historical pattern needs different evidence from one meant to explore what might happen. Purpose comes first in the ODD protocol (Overview, Design concepts, Details), a standard template for describing ABMs whose second update Volker Grimm and colleagues published in 2020.
2. Could you rebuild it from the paper? ODD asks for the entities and their variables, the scales, the schedule, the starting conditions, the input data and every submodel. If you could not code it from the description, neither could a skeptic.
3. How many patterns does it match? Grimm and colleagues' pattern-oriented modeling (2005, Science) argues that a model should reproduce several patterns at different scales at once, say a population curve, a settlement map and an age structure, because fitting one curve is easy.
4. What happens when the parameters move? If small changes flip the result, be wary. Janssen's reanalysis of Long House Valley is a good example of the check.
5. Would something simpler do as well? If a carrying-capacity sum or a two-line recursion gives the same answer, the agents may not be earning their place.
6. How many runs, and how wide was the spread? Our six households showed how far a single run can mislead.
7. Where did the rules come from? From ethnography, from experiments, or for convenience? A rule such as sharing at 4 units or more is a hypothesis about how people behave, and it deserves the same scrutiny as any other claim about a living community.

Some tools on this site are rule-based in a simpler way: the section system simulator follows marriage and descent rules rather than individual choices, and the others are listed on the models page.

Sources and further reading

• Epstein, Joshua M. 1999. "Agent-based computational models and generative social science." Complexity 4 (5): 41-60.
• Axtell, Robert L., Joshua M. Epstein, Jeffrey S. Dean, George J. Gumerman, Alan C. Swedlund, and colleagues. 2002. "Population growth and collapse in a multiagent model of the Kayenta Anasazi in Long House Valley." Proceedings of the National Academy of Sciences 99 (suppl. 3): 7275-7279. https://doi.org/10.1073/pnas.092080799
• Janssen, Marco A. 2009. "Understanding Artificial Anasazi." Journal of Artificial Societies and Social Simulation 12 (4): 13. https://www.jasss.org/12/4/13.html
• Lansing, J. Stephen, and James N. Kremer. 1993. "Emergent properties of Balinese water temple networks: coadaptation on a rugged fitness landscape." American Anthropologist 95 (1): 97-114.
• Kuznar, Lawrence A., and Robert Sedlmeyer. 2005. "Collective violence in Darfur: an agent-based model of pastoral nomad/sedentary peasant interaction." Mathematical Anthropology and Cultural Theory 1 (4). https://escholarship.org/uc/item/67x4t8ts
• Grimm, Volker, and colleagues. 2005. "Pattern-oriented modeling of agent-based complex systems: lessons from ecology." Science 310 (5750): 987-991.
• Grimm, Volker, Steven F. Railsback, Christian E. Vincenot, and colleagues. 2020. "The ODD protocol for describing agent-based and other simulation models: a second update to improve clarity, replication, and structural realism." Journal of Artificial Societies and Social Simulation 23 (2): 7. https://www.jasss.org/23/2/7.html