This is the supporting code for:

Greg d'Eon, Hala Murad, Kevin Leyton-Brown, and James R. Wright. ElementaryNet: A Non-Strategic Neural Network for Predicting Human Behavior in Normal-Form Games. AAAI 2026 (to appear). arXiv:2503.05925.

In a virtual environment,

pip install -e .
pip install -r requirements.txt

Some scripts assume that the BGT_BASE_DIR environment variable points to this folder: e.g.,

export BGT_BASE_DIR="path/to/this/folder"

In the scripts directory, python commands.py generates a list of commands that will reproduce the results from the paper. These commands are split into separate experiments, which each train the following models, respectively:

• qch: baseline Uniform + QCHp model
• gamenet: GameNet + QCHp models
• enet-qchp: ElementaryNet + QCHp models with learned potentials (including the best-performing ElementaryNet model)
• enet-level0: purely level-0 ElementaryNet models with no strategic model
• enet-qchk: ElementaryNet + QCH{1, 2, 3} models
• enet-own: ElementaryNet + QCHp models with one potential set to the "own" function
• enet-fixed: ElementaryNet + QCHp models with all four fixed potentials

python plot.py generates the figures.

Files in the data folder are taken from the following sources:

• (Wright and Leyton-Brown, 2019): .nfg files containing games in all10 dataset
• (Fudenberg and Liang, 2019): .mat files containing both games and observations
• (Chui, Hartline, and Wright, 2023): .nfg files containing games and .pkl files containing observations