SWM-AP

The code for paper Social World Model-Augmented Mechanism Design Policy Learning.

SWM-AP augments mechanism-design policy learning with a learned social world model. This repository keeps the official SWM-AP reproduction path for three environments: Facility Location, AdaSociety, and AI-Economist.

The official launcher carries the default replication settings internally.

# Facility Location
ENVIRONMENT=facility scripts/launch_official_repro.sh

# AdaSociety
ENVIRONMENT=adasociety scripts/launch_official_repro.sh

# AI-Economist
ENVIRONMENT=aieconomist scripts/launch_official_repro.sh

More details are in docs/EXPERIMENTS.md.

SWM_AP/
  algorithms/        Facility Location SWM-AP code and small checkpoints
  experiments/       AdaSociety and AI-Economist SWM-AP entrypoints
  AdaSociety/        vendored AdaSociety environment
  ai-economist/      vendored AI-Economist environment
  config/            AI-Economist experiment configuration
  docs/              experiment guide and reproducibility notes
  paper_results/     compact result snapshots
  requirements/      dependency lists by environment
  scripts/           official launcher and TensorBoard extraction tools

Baseline training implementations are intentionally not included in this public release branch for now. The committed CSVs keep compact comparison summaries so the paper-facing deltas remain auditable.

Compact, anonymized result summaries are stored under paper_results/ and mapped in docs/PAPER_RESULTS_MANIFEST.md. Full TensorBoard logs are intentionally not stored in git.

Install only the environment stack you need:

python -m pip install -r requirements/facility.txt
python -m pip install -r requirements/adasociety.txt
python -m pip install -r requirements/aieconomist.txt

For AI-Economist, Python 3.8 is recommended. Install the vendored package in editable mode:

python -m pip install -e ai-economist --no-deps

python scripts/extract_tensorboard_scalars_stream.py \
  --runs-root <RUNS_ROOT> \
  --tag <TENSORBOARD_TAG> \
  --output results/scalars.csv

python scripts/summarize_scalar_runs.py \
  --input results/scalars.csv \
  --output results/summary.csv \
  --aggregate-output results/aggregate.csv \
  --tag <TENSORBOARD_TAG>