BenchMFG

Benchmark suite for Mean Field Game algorithms.

Contents

• Install
• Quick Start
• Registered Configs: envs and algos
• Sweep
• Python API
• Outputs And Plots
• Extending BenchMFG: create new envs and algos
• MF-Garnet: scaling laws and robustness

Install

uv add bench-mfg-suite
# or
pip install bench-mfg-suite

For local development:

uv sync --extra dev

CUDA is optional (but if we are running on gpu make sure you download that!!!!). The default install uses CPU-compatible JAX.

# Linux/NVIDIA, pip-managed CUDA runtime components:
uv add "bench-mfg-suite[cuda12]"
pip install "bench-mfg-suite[cuda12]"

# Linux/NVIDIA, local CUDA installation:
pip install "bench-mfg-suite[cuda12-local]"

If GPU initialization fails, check nvidia-smi and the official JAX install matrix: https://docs.jax.dev/en/latest/installation.html

Quick Start ⚡️

👑 Quick understanding of the repository and the package run the following three commands

benchmfg hello
benchmfg garnet
benchmfg mfpso

List registered configs:

benchmfg env list
benchmfg algo list

💥 Run one quick experiment on your machine:

benchmfg train algorithm=omd environment=four_rooms_obstacles device=cpu

Registered Configs

Environments: contraction_game, four_rooms_obstacles, kinetic_congestion, lasry_lions_chain, mf_garnet, multiple_equilibria, no_interaction_game, potential_game2d, rock_paper_scissors, sis_epidemic.

Algorithms: damped_fixed_point, omd, pi, pso.

Use benchmfg env list and benchmfg algo list for the installed package’s authoritative list.

Use benchmfg algo-parameters (or make algo-parameters) to print every algorithm’s hyperparameters, defaults, recommended sweep ranges, and the exact override syntax.

Sweep

Run a sweep:

benchmfg sweep \
  algorithm=omd \
  environment=lasry_lions_chain \
  experiment.name=omd_sweep \
  experiment.random_seed=42,10,111,1032 \
  algorithm.omd.learning_rate=0.5,0.05,0.005 \
  algorithm.omd.temperature=0.2,0.5,0.8

Python API

import benchmfg

cfg = benchmfg.load_config(["algorithm=omd", "environment=lasry_lions_chain"])
environment, initial_policy = benchmfg.make_environment(cfg)
solver = benchmfg.make_solver(
    cfg,
    environment=environment,
    initial_policy=initial_policy,
)

fixed_mf_env = benchmfg.make_fixed_mean_field_env(
    environment,
    environment.stationary_mean_field,
)

Outputs And Plots

Runs write artifacts under:

outputs/<Env>/<Algorithm>/seed_<seed>/<Experiment>/<run_id>/

Important files: exploitabilities.npz, final_mean_field.npz, final_policy.npz, metrics.npz, config.yaml.

Plot commands:

benchmfg plot single-run <run_dir>
benchmfg plot sweep <environment> <algorithm>
benchmfg plot compare <environment>

Plot discovery defaults:

• single-run <run_dir> plots exactly that timestamped run.
• sweep <environment> <algorithm> scans outputs/ by default. For each seed and hyperparameter version, it selects the latest timestamped run containing exploitabilities.npz.
• compare <environment> reads the results/<environment>/<algorithm>/best_model.yaml files written by plot sweep; rerun plot sweep first if new runs were added.
• Use --outputs-dir <path> on sweep/compare commands when artifacts are not under outputs/.

Repository Layout

src/benchmfg/
├── config/      # packaged Hydra configs
├── envs/        # MFG environments
├── learner/     # solvers
├── utility/     # training, saving, plotting helpers
├── cli.py       # benchmfg command
└── train.py     # Hydra train entrypoint

See EXPERIMENTS.md for batch-run scripts.