Simulation Paradigms
Agent-based models, cellular automata, and feedback systems in geographic modelling
Before You Start
You should know
That some geographic systems evolve through time and that the order of events can matter as much as the starting conditions.
You will learn
How three major simulation paradigms differ, what questions each one answers best, and why simulation is useful when static equations or one-shot spatial operations are not enough.
Why this matters
Many important geographic problems are fundamentally about process over time: spread, growth, routing, adaptation, and feedback.
If this gets hard, focus on…
The basic distinction: are we updating cells, individual agents, or aggregate stocks and feedback loops?
Some geographic questions cannot be answered by one equation or one overlay operation because the path is the point. A wildfire moves cell by cell across fuel and wind fields. Urban expansion depends on local neighborhoods, accessibility, and contingent development sequences. A migration system depends on the decisions of many individuals who react to one another and to changing opportunity. In each case, the final pattern emerges from repeated updates through time. That is the domain of simulation.
But “simulation” is not one method. It is a family of methods. Some simulations update a grid of cells according to local transition rules. Some update individual agents with their own states and decision rules. Some track aggregate stocks and flows linked by reinforcing or balancing feedbacks. These approaches can look similar from a distance because they all produce time-evolving outputs, but they are best for different kinds of structure.
1. The Question
When should we switch from a static model to a simulation?
Usually when one or more of these is true:
- the order of events matters
- local interactions create larger-scale patterns
- feedback loops change the future state of the system
- a single final answer is less useful than a time-evolving scenario
The next decision is which style of simulation fits the structure of the problem.
2. The Conceptual Model
Simulation Style Should Match What Is Actually Changing Through Time
The cleanest distinction is structural. If the evolving thing is a grid, use cell-based rules. If it is many individuals, use agents. If it is a few coupled totals with feedback, use a stock-and-flow style model.
Best For Spatial Spread On A Grid
Each cell updates according to its own state, its neighbors, and maybe an external forcing field. Good for wildfire, land change, contagion surfaces, and neighborhood effects.
Best For Many Heterogeneous Decision-Makers
Each agent has its own rules, memory, and movement or choice behavior. Good for mobility, migration, land-use actors, and systems where micro-decisions generate macro-patterns.
Best For Aggregate Feedback Loops
Track totals, rates, delays, and feedbacks rather than individual cells or actors. Good for reservoirs, resource depletion, population sectors, and policy feedback systems.
1. Cellular automata
A cellular automaton updates a spatial grid one step at a time. Each cell changes state according to a rule that depends on:
- its current state
- the states of neighboring cells
- optional external conditions like wind, slope, or accessibility
This is often the right choice when space is explicitly gridded and local adjacency is the main engine of change.
Typical examples:
- fire spread
- land-use change
- simple flood expansion
- ecological patch dynamics
Strength:
- preserves local spatial contagion or spread very naturally
Weakness:
- individual decisions and long-range adaptation are often oversimplified
2. Agent-based models
An agent-based model (ABM) simulates many individual entities:
- households
- firms
- migrants
- vehicles
- animals
Each agent may have its own state, goals, resources, and decision rules. Agents interact with one another and with a landscape.
This is often the right choice when heterogeneity matters and when aggregate behavior emerges from many decentralized actions.
Strength:
- captures heterogeneous actors and emergent patterns
Weakness:
- can become difficult to calibrate, explain, and validate
3. System dynamics and stock-flow models
This family tracks aggregate quantities and the flows between them:
- water stored in a reservoir
- population in age classes
- carbon in atmosphere, biomass, and soil
- capital, demand, and infrastructure capacity
The central idea is feedback:
- reinforcing loops amplify change
- balancing loops resist change
- delays create overshoot and oscillation
Strength:
- excellent for policy reasoning and feedback-rich systems
Weakness:
- can hide important spatial detail or heterogeneity
3. A Decision Workflow
Use this sequence:
- If adjacency on a raster or lattice is the main engine, start with a cellular automaton.
- If different actors make different decisions, start with an agent-based model.
- If the question is mainly about coupled totals and feedback loops, start with a stock-flow or system-dynamics model.
A few examples
| Question | Better simulation backbone | Why |
|---|---|---|
| How does fire spread over changing fuels and wind? | cellular automaton | local spread on a grid is the central process |
| How do households choose locations as a city expands? | agent-based model | heterogeneous choices matter |
| How do storage, inflow, release, and demand interact in a reservoir? | stock-flow model | aggregate feedbacks dominate |
Hybrids are common
Strong simulation projects often mix paradigms:
- cell-based fire spread + probabilistic weather ensembles
- agent decisions on top of a GIS suitability surface
- stock-flow water budgeting + spatial routing
As with model choice more generally, the first task is picking the right backbone.
4. Worked Example by Hand
Imagine three versions of urban growth modelling for the same region.
Version A: cellular automaton
Each grid cell is either urban or non-urban. At each time step, a non-urban cell becomes urban if:
- enough neighbors are already urban
- accessibility is high
- slope is low
This is good if the main interest is the geometry of urban expansion.
Version B: agent-based model
Households and developers each choose among locations based on:
- price
- travel time
- amenity
- past occupancy
This is better if the interest is competition, behavior, and heterogeneity.
Version C: stock-flow model
Track:
- total housing stock
- total demand
- construction rate
- infrastructure capacity
This is better if the interest is feedback between aggregate demand, supply, and policy.
The lesson is not that one is always right. The lesson is that each one answers a different urban-growth question.
5. What Could Go Wrong?
Using a cellular automaton for a system driven by strategic actors
If the real story is heterogeneous decision-making, neighbor rules alone may flatten the important behavior.
Building an ABM when the real need is only aggregate feedback
An agent-based model can add heavy complexity where a simpler stock-flow model would be clearer and more testable.
Treating system-dynamics models as if space does not matter
If location is central to the mechanism, an aggregate model may miss the whole geography of the problem.
Forgetting validation
Simulation outputs can look persuasive because they are dynamic and visual. They still need calibration, out-of-sample checks, and uncertainty analysis.
Summary
- Simulation is not one thing; it is a family of time-evolving model structures.
- Cellular automata are strongest for local spread on grids.
- Agent-based models are strongest for heterogeneous actors and emergent patterns.
- System-dynamics models are strongest for aggregate feedback loops and policy scenarios.
- The best paradigm is the one that matches what is actually changing through time.