Agent-Based Modelling of Social Systems

The appeal of agent-based modelling (ABM) is that it lets you state your assumptions at the level where you actually have knowledge. You may have no idea what the aggregate demand curve for a neighbourhood looks like, but you may know how households decide whether to move, whom they talk to, and what they can afford. ABM turns micro-level rules into macro-level consequences, and the gap between the two is where the interesting results live.

The canonical demonstration

Schelling's segregation model remains the clearest teaching example, and it is worth stating precisely because it is so often misdescribed. Agents live on a grid. Each has a mild preference: it wants at least some fraction of its neighbours to be like itself. Nobody wants segregation. Run the model and you get segregation anyway, sharply, from preferences that would be considered tolerant by any reasonable standard. The macro pattern is not in the micro rule. That gap is the whole subject.

What ABM is good at

• Heterogeneity that matters. When the variance across agents drives the outcome rather than the mean.
• Interaction structure. When who talks to whom changes the result - see social networks.
• Adaptation. When agents change their rules in response to what happened last period.
• Counterfactuals that cannot be observed. Policy variants that were never run in the real world.

What ABM is bad at

Prediction of specific quantities at specific dates. Any model flexible enough to represent adaptive heterogeneous agents is flexible enough to fit almost anything, which is exactly why validation and calibration is the part of the workflow that separates useful models from expensive animations. Start with how a model is actually put together, then look at the software landscape.