AgentDOOM is a reinforcement learning research project focused on training PPO (Proximal Policy Optimization) agents to play Atari Breakout, comparing different visual encoders such as CNNs and Vision Transformers (ViTs).
- Actor-Critic architecture using PPO
- Visual state encoding with CNN and ViT backbones
- Experimental setup for encoder comparison in Atari environments
AgentDOOM/
├── agents/ # CNN and ViT agent definitions
├── training/ # Training and evaluation scripts
├── notebooks/ # Experimental notebooks
├── models/ # Saved checkpoints
├── logs/ # Training logs and metrics
├── gifs/ # Gameplay visualizations
├── requirements.txt
├── pyproject.toml
└── README.md