32  Exploring parameter space

In the previous chapter, we used BehaviorSpace to run structured simulation experiments. In this chapter, we step back and ask a broader methodological question: how should we explore a model’s parameter space, and why?

Agent-based models such as Artificial Anasazi typically contain multiple parameters that are:

Parameter exploration is therefore not about “finding the right value”, but about understanding the space of plausible behaviours the model can generate.

32.1 What is parameter space?

Each parameter in a model defines one dimension of a multidimensional space. A single simulation run corresponds to one point in that space.

In the Artificial Anasazi model, parameters such as:

  • harvest-related parameters
  • fertility or death-related parameters
  • maize productivity or environmental constraints (data input)

together define a parameter space in which different combinations may lead to:

  • long-term persistence,
  • oscillating population sizes,
  • rapid collapse and abandonment.

Even with only a handful of parameters, the number of possible combinations grows very quickly.

NoteArtificial Anasazi example

If we explore only three parameters, each with 10 possible values, this already produces 10 × 10 × 10 = 1,000 configurations, before accounting for repetitions.

32.2 Why not explore everything?

An exhaustive exploration of parameter space is almost never feasible:

  • the number of combinations explodes combinatorially,
  • many combinations are implausible,
  • simulation time and data storage become limiting factors.

More importantly, exhaustive exploration is rarely conceptually useful. The goal is not to catalogue all possible outcomes, but to identify:

  • sensitive parameters,
  • thresholds and regime shifts,
  • regions of stability and instability.

This motivates different sampling strategies.

32.3 Structured parameter exploration (grids and sweeps)

BehaviorSpace is particularly well suited for structured exploration, where parameters are varied systematically across fixed values or ranges.

Typical strategies include:

  • varying one parameter while keeping others fixed,
  • exploring small grids of two or three parameters,
  • increasing resolution around interesting regions.

In the Artificial Anasazi model, this might involve:

  • sweeping harvest-variance from low to high values,
  • comparing population trajectories across these levels,
  • identifying values beyond which collapse becomes common.
NoteArtificial Anasazi example

A one-dimensional sweep of harvest-variance allows us to ask: How much environmental unpredictability can the system tolerate before settlement patterns break down?

Structured exploration is intuitive and easy to interpret, but it scales poorly as the number of parameters increases.

32.4 Alternative designs or input data

Some times existing alternative design or “competing” versions of the model can be worth exploring. for this we can use the mediation of nominal parameters, which in NetLogo would typically reside in the interface as switches and choosers widgets. This is not available in the Artificial Anasazi, but the chooser simulation-period in the Mesara Trade could be considered in this way.

32.5 Random and quasi-random sampling

When the parameter space becomes too large for structured sweeps, an alternative approach is to sample it.

Instead of testing all combinations, we:

  • define plausible ranges for each parameter,
  • draw parameter values randomly (or quasi-randomly),
  • run the model for each sampled configuration.

This approach treats the model as a generator of outcomes, which we probe statistically.

Random sampling is particularly useful when:

  • interactions between parameters are expected,
  • we are interested in global patterns rather than fine-grained tuning,
  • we plan to analyse results in R.
NoteArtificial Anasazi example

Rather than varying only harvest-variance, we might simultaneously sample:

  • harvest-adjustment (the fraction of the harvest that is actually usable)
  • fertility-related parameters
  • mortality thresholds

to ask which combinations tend to produce persistence versus collapse.

32.6 Defining plausible parameter ranges

Before sampling, parameter ranges must be defined carefully.

For archaeological models, this step is theoretical, not technical. Ranges should reflect:

  • archaeological estimates,
  • ethnographic analogies,
  • exploratory uncertainty rather than extremes.

For example, in the Artificial Anasazi model:

  • harvest-variance might reasonably vary between low and moderate unpredictability,
  • demographic parameters should avoid biologically implausible values,
  • environmental productivity should reflect the known limits of maize agriculture.
WarningImportant

Wide parameter ranges increase coverage, but also increase the risk of generating meaningless or misleading outcomes.

32.7 Systematic random sampling in R

Once parameter ranges are defined, R can be used to generate a design of experiments that is then passed to NetLogo.

A simple random sampling workflow in R involves:

  1. defining parameter ranges,
  2. drawing random samples,
  3. exporting them as a table,
  4. running NetLogo for each row.

32.7.1 Conceptual example (R)

set.seed(123)

n <- 200  # number of sampled configurations

params <- data.frame(
  harvest_variance = runif(n, min = 0.2, max = 0.8),
  fertility_rate   = runif(n, min = 0.02, max = 0.06),
  death_age        = sample(35:45, n, replace = TRUE)
)

Each row of this table represents one plausible Anasazi world.

These configurations can then be:

  • read from file during BehaviorSpace setups,
  • read from a file in a specific procedure in NetLogo, or
  • looped over using RNetLogo or nlrx, to execute simulations from R.

32.7.2 Interpreting sampled results

Unlike structured sweeps, random sampling does not produce neatly ordered results. Interpretation therefore, relies on:

  • visualisation (scatter plots, density plots),
  • summary statistics,
  • comparisons between outcome categories (e.g. persistence vs. collapse).

For the Artificial Anasazi model, typical analyses might include:

  • relating the final population size to sampled parameters,
  • identifying parameter regions associated with early abandonment,
  • comparing distributions of outcomes rather than single trajectories.
NoteArtificial Anasazi example

Instead of asking “What happens when harvest variability is 0.6?”, we ask “Under which combinations of parameters does collapse become likely?”

32.7.3 Complementarity with BehaviorSpace

BehaviorSpace and R-based sampling are not competing approaches.

  • BehaviorSpace excels at structured, transparent experiments.
  • R-based sampling excels at large-scale, flexible exploration.

In practice, a common workflow is:

  1. structured exploration to build intuition,
  2. random sampling to explore interactions,
  3. refined sweeps around interesting regions.

This iterative process reflects how agent-based models are used as theoretical laboratories, not predictive machines.

32.8 Exercises

TipExercise 1: Conceptual

List three parameters of the Artificial Anasazi model.

  • Which are well-constrained archaeologically?
  • Which are highly uncertain?
  • Which would you prioritise for exploration, and why?
TipExercise 2: Design

Define plausible ranges for two parameters of the model.

  • Justify your choices in archaeological terms.
  • Discuss what kinds of outcomes you expect at the extremes.
TipExercise 3: Sampling logic

Assume you can only afford 200 simulation runs.

  • Would you prefer a structured grid or random sampling?
  • What questions would each approach allow (or prevent) you from answering?