Predator-Prey-Grass

Multi-Agent Deep Reinforcement Learning meets Darwinian and Baldwinian evolution

A multi-agent reinforcement-learning ecosystem for studying how cooperation emerges through learning within lifetimes and evolution across generations.

_{Fixed-trait game-theoretic hunting: coevolution, cooperation, defection and free-riding emerging under a fixed reward design.}

• Darwinian evolution of inherited traits (speed, cooperation rate, metabolic rate, and more) via reproduction and mutation
• Baldwinian interaction between evolution and learning tested directly — do genetic selection and learned behavior actually shape each other, or just coexist?
• Emergent cooperation, defection, reciprocity and coevolution, studied under both evolving and fixed-trait agent populations
• Built with Python, Gymnasium and RLlib 2.59's new API stack (RLModule / Learner / EnvRunner), with dynamic, lifecycle-changing agent populations

Start here: Quick start to run a demo in under five minutes, or EXPERIMENTS.md for the full catalogue of experiments.

How it works

This project explores whether cooperative behavior, coevolution, defection, and free-riding can emerge and stabilize in a spatial, resource-limited ecosystem, by combining within-lifetime multi-agent reinforcement learning with population-level ecological and evolutionary dynamics. It probes the interplay between nature (inherited traits via reproduction and mutation) and nurture (behavior learned via reinforcement learning) — including a direct test of the Baldwin effect: whether genetic selection and learned behavior actually shape each other, not just coexist. Agents differ by speed, vision, energy metabolism, and decision policies — offering ground for open-ended adaptation. At its core lies a gridworld simulation where agents are not just trained — they are born, age, reproduce, die, and even mutate in a continuously changing environment.

Legacy snapshot: the pre-cleanup research codebase is archived at PredPreyGrassLegacy.

Start here

The repo splits into two structurally different families of experiment, matching the predpreygrass/evolutionary/ vs predpreygrass/non_evolutionary/ directory split:

• Evolutionary: agents carry a heritable genome trait, passed parent → offspring with mutation. What gets selected is discovered, not designed.
• Non-evolutionary: every agent trait is fixed; only the RL policy adapts. What emerges is a behavioral equilibrium under a given incentive design, not a change in the population's genetics.

The full catalogue of environments and experiments — evolutionary trials, reward-shaping variants, cooperation/game-theory environments, and the Red Queen evaluations — lives in EXPERIMENTS.md.

Quick start: run a demo in under five minutes

git clone https://github.com/doesburg11/PredPreyGrass.git
cd PredPreyGrass
python -m venv .venv && source .venv/bin/activate
pip install -e .
python ./predpreygrass/non_evolutionary/discrete_space/base_environment_discrete/random_policy.py

This installs the base dependencies and runs a random policy in the base environment — no VS Code or Conda required. pygame (for the rendered window) ships as a regular pip dependency; the Conda/GCC setup below is only needed if your platform lacks a prebuilt pygame wheel.

Pretrained checkpoints and historical training outputs are preserved in the legacy archive rather than shipped in the active source tree.

Full setup (Visual Studio Code + Conda)

Editor used: Visual Studio Code on Linux Mint 22.0 Cinnamon

1. Clone the repository:

   git clone https://github.com/doesburg11/PredPreyGrass.git

2. Open Visual Studio Code and execute:
   • Press ctrl+shift+p
   • Type and choose: "Python: Create Environment..."
   • Choose environment: Conda
   • Choose interpreter: Python 3.11.13 or higher
   • Open a new terminal

   • pip install -e .

3. Install the additional system dependency for Pygame visualization (only needed if pip install can't find a prebuilt pygame wheel for your platform):

   • conda install -y -c conda-forge gcc=14.2.0

Acknowledgments

Developed with AI coding assistance from Claude (Anthropic), which does the implementation, with Codex (OpenAI) acting as an independent second opinion, peer-reviewing Claude's nontrivial code changes.

References

• RLlib: Industry-Grade, Scalable Reinforcement Learning
• Paper Collection of Multi-Agent Reinforcement Learning (MARL)
• Multi-Agent Reinforcement Learning: Foundations and Modern Approaches. Stefano V. Albrecht, Filippos Christianos, and Lukas Schäfer

Citation

If you use this software in your research, please cite it — see CITATION.cff.