RL Multi-agent Boxing Project

A reinforcement learning project built for the second CM50270 coursework at the University of Bath.

This project trains agents (like PPO, DQN, and others) to play Atari Boxing using the Gym ALE environment.

---

Getting Started

Windows users:
The multi-agent-ale-py package does not compile on native Windows.
You must use Windows Subsystem for Linux (WSL) to run this project.
Once WSL is installed, follow the WSL/Linux instructions below.

---

1. Clone the repository

git clone https://github.com/YOUR-TEAM-NAME/rl-boxing.git
cd rl-boxing

---

2. Create a virtual environment (recommended)

python3 -m venv venv

# For Mac/Linux/WSL:
source venv/bin/activate

---

3. Install Dependencies

If you're rendering the boxing gym locally:

pip install -r render_requirements.txt

If you're rendering in WSL (with X server like VcXsrv):

sudo apt update
sudo apt install cmake swig zlib1g-dev libboost-all-dev \
                 libsdl2-dev libsdl2-image-dev \
                 python3-dev build-essential

pip install -r og_requirements.txt

# Then add this to ~/.bashrc or ~/.zshrc:
export DISPLAY=:0.0

If you're training agents (on Hex):

pip install -r (your requirements txt file name).txt

---

4. Download Atari ROMs

AutoROM --accept-license

---

5. Run a test match (e.g., RandoAgent1 vs RandoAgent2)

python main_rando.py

---

6. Train PPO or DQN agents

python training/train_ppo.py
python training/train_dqn.py

---

7. Watch trained agents compete

python main.py

---

Agents

• RandoAgent1: Random agent with equal probabilities
• RandoAgent2: Slower/weaker random agent (for testing)
• PolicyAgent: PPO agent (trained via train_ppo.py)
• DQNAgent: DQN agent (trained via train_dqn.py)

---

Project Structure

rl-multiagent-boxing/
├── agents/
│   ├── __init__.py
│   ├── dqn_agent.py
│   ├── policy_agent.py
│   ├── rando_agent1.py
│   └── rando_agent2.py
├── models/
│   ├── ppo_model.h5
│   └── dqn_model. (keras or h5)
├── training/
│   ├── train_ppo.py
│   └── train_dqn.py
├── main.py                                 # PPO vs DQN match
├── main_rando.py                           # RandoAgent1 vs RandoAgent2
├── training_requirements_draft.txt         # A SAMPLE FILE FOR TRAINING ON HEX
├── rendering_requirements.txt              # For local rendering
├── og_rendering_requirements.txt           # For WSL local rendering
└── README.md

---

References

• Gymnasium ALE Environments
• AutoROM tool

---

University of Bath — CM50270 Reinforcement Learning Coursework 2