Agentic Uplift

Multi-agent reinforcement learning (MARL) framework with an electricity market example. Includes PPO-based agents, Gymnasium-compatible environments, and reproducible sequential/parallel training.

Install

Requires Python 3.11+ and uv.

uv sync --extra dev

Editable install is set up automatically from pyproject.toml.

Project Layout

easy_marl/
  src/                    # Core MARL framework
    agents.py             # Base + PPO agents
    environment.py        # Gymnasium-compatible MARL env
    observators.py        # Observation builders
  examples/
    bidding/              # Electricity market example
      market.py           # Market clearing + utils
      training.py         # Sequential/parallel training
tests/                    # Unit tests
outputs/                  # Training/eval artifacts (ignored)
main.py                   # Simple entry point
pyproject.toml            # Build + deps

Quickstart

Run the default training:

uv run python main.py

Use the bidding example training directly:

uv run python -c "from easy_marl.examples.bidding.training import parallel_train; parallel_train(N=3, num_rounds=1, timesteps_per_agent=48)"

Tests

uv run pytest -q

Key Concepts

• MARL environment returns agent-specific observations built via observers.
• Two training regimes:
  • Sequential (Iterated Best Response / IBR)
  • Parallel (Simultaneous Best Response / SBR)
• Reproducibility via explicit seeding (NumPy/Torch/processes).

Development Notes

• Package installs as easy_marl (editable).
• Example code lives under easy_marl/examples/bidding to keep core MARL generic.
• Outputs are written to outputs/ and excluded from packaging.

License

Choose and add a license (e.g., MIT or Apache-2.0) before public release.