simulatingrisk 1.0.0.post1

Agent-based modeling for simulations related to risk and rationality

Simulating Risk

The code in this repository is associated with the CDH project Simulating risk, risking simulations.

Simulations are implemented with Mesa, using Agent Based Modeling to explore risk attitudes within populations.

---

Simulations with agent interaction

• Hawk/Dove with risk attitudes
• Hawk/Dove with multiple risk attitudes and adjustment

The code for Hawk/Dove with risk attitudes and Hawk/Dove with multiple risk attitudes in this codebase was reviewed in June 2024 by Scott Foster and Malte Vogl (Senior Research Fellow, Max Planck Institute of Geoanthropology) via DHTech Community Code Review; review was faciliated by Cole Crawford (Senior Software Engineer, Harvard Arts and Humanities Research Computing).

View an interactive version of the Hawk/Dove with multiple risk attitudes simulation online.

The simulation can also be run in a Jupyter notebook:

Simulations with risky choices (environment)

• Risky Food
• Risky Bet

Risk attitude definitions

Across simulations, we define agents with risk attitudes tracked via a numeric r or risk_level 0.0 - 1.0, where r is that agent's minimum acceptable risk level for taking the risky bet. When the probability 'p' of the risky bet paying off is greater than an agent's 'r', that agent will take the bet. An agent with r=1 will always take the safe option (no risk is acceptable); an agent with r=0 will always take the risky choice (any risk is acceptable). Notice that the agent is never indifferent; allowing indifference would require introducing a tie-breaking rule, which would be a further parameter.

When the risky bet might be better or worse than the safe bet by the same amount (for example, the risky bet yields 3 or 1 and the safe bet yields 2), an agent who maximizes expected utility will prefer the risky bet when p > 0.5 and will prefer the safe bet when 'p < 0.5'; and they will be indifferent between the risky bet and the safe bet. Thus, r = 0.5 corresponds to expected utility maximization except in the case in which the probability is exactly 0.5 (or, we might say, a point epsilon units to the left of 0.5, where epsilon is smaller than the fineness of our random number generator, corresponds to expected utility maximization). These complications make no difference in practice, so we can simply say that r = 0.5 corresponds to expected utility maximization.

Development instructions

Initial setup and installation:

• Recommmended: create and activate a Python 3.12 virtualenv:

python3 -m venv simrisk
source simrisk/bin/activate

• Install the package, dependencies, and development dependencies:

pip install -e .
pip install -e ".[dev]"

To run the Solara app with tabs for all simulations that are available as Solara apps:

solara run simulatingrisk/app.py

To run a single simulation, run Solara with the model-specific app file:

solara run simulatingrisk/hawkdovemulti/app.py

An interactive version of the Solara app with tabs for all simulations is available online.

Running the simulations (non-solara versions)

• Simulations can be run interactively with mesa runserver by specifying the path to the model, e.g. mesa runserver simulatingrisk/risky_bet/ Refer to the readme for each model for more details.
• Simulations can be run in batches to aggregate data across multiple runs and different parameters. For example, ./simulatingrisk/batch_run.py riskyfood

Install pre-commit hooks

Install pre-commit hooks (currently black and ruff):

pre-commit install