Agent-based modeling
Model individual people or objects when their differences, interactions, or locations change the result.
An agent is an individual actor with its own state and behavior. It may represent a person, vehicle, organization, animal, device, or facility. Agents can differ from one another and change over time.
Each agent follows local rules, yet their interactions can produce a larger pattern: congestion, adoption, cooperation, an epidemic, or a market shift. No agent needs to direct the overall result.
Model individuals only when their differences or interactions affect the result.
- state
- active
- location
- (12, 7)
- rule
- observe → decide → act
For example: clustering, traffic flow, adoption, or crowd movement.
Use an agent-based model when individual differences, relationships, movement, or adaptation can change the outcome.
- Individuals have different attributes or strategies.
- Who interacts with whom matters.
- Spatial location or a network structure matters.
- Behavior changes in response to experience or other agents.
An agent model can be sensitive to who exists at time zero, where agents are placed, and which relationships already exist. Match those starting distributions to the study population when data are available, rather than scattering agents uniformly because it is convenient.
If the initial network or spatial pattern is uncertain, vary it across replications or scenarios and check whether the conclusion changes.
Verify that one agent behaves correctly in controlled cases, then compare aggregate patterns such as adoption curves, contact rates, congestion, or spatial clustering with observations the model is expected to reproduce.
A convincing population-level curve does not prove the individual rules are right; different micro-rules can sometimes produce similar aggregate output.
Detailed agents require more data, assumptions, and computing time. If identity, interaction, or location does not affect the decision, a process or aggregate model is easier to build, explain, and test.