Cooperation or Defection: Multi-Agent Learning in a 3D Continuous Social Sequential Dilemma Game

This project explores cooperative and competitive behaviors between predator agents (wolves) chasing a prey agent (goat) within a realistic Unity environment featuring obstacles and raycast-based vision.

Project Overview

The WolfpackUnity ML-Agents project examines how agents learn and adapt behaviors of cooperation or defection through multi-agent reinforcement learning techniques. The goal is to analyze agent interactions in dynamic scenarios, providing insights into cooperative strategies and emergent behaviors.

Project Structure

• Assets/

  • Scenes/: Contains two primary Unity scenes:
    • train: For training agents.
    • inference: For visualizing trained agents' behaviors.
  • Scripts/: Includes core C# scripts controlling agents and environment interactions:
    • PreyAgent.cs
    • PredatorAgent.cs
    • EnvironmentController.cs
    • CameraView.cs
    • Radius.cs
  • ML-Agents/: Contains trained neural network weights for agents.
  • Unity skins and prefabs for visual representation.

• Packages/: Unity-managed packages required for the environment.

• ProjectSettings/: Unity editor configurations and settings.

• config/: YAML configuration files for training agent behaviors using ML-Agents.

• .collabignore: Configuration for Unity Collaborate (ignores unnecessary files).

• update_config.py: Python script to streamline YAML configuration updates from the command line for efficient experimentation.

Installation & Setup

Prerequisites

• Unity Editor version 6000.1.1f1
• ML Agents Package 3.0.0
• Python installed with dependencies listed in requirements.txt

Quick Start

1. Clone this repository.
2. Install Python dependencies:

   pip install -r requirements.txt

3. Open the Unity project located in the WolfpackUnity folder using Unity Editor.
4. Train agents using ML-Agents:

   mlagents-learn config/your_config.yaml --env your_train_env_build --run-id run_identifier --no-graphics

5. Once trained, run inference scene in Unity to observe agent behaviors.

Usage

• Modify training parameters quickly by adjusting your YAML file or using the provided update_config.py:

  python update_config.py config/base_config.yaml config/new_config.yaml --param param_1_name=param_1_value --param param_2_name=param_2_value

• Experiment with different scenarios and observe cooperation or defection patterns between predator agents.

Example