How ulrikisdahl/muzero-breakout works
Project type and purpose
• This is a research-style reinforcement learning (RL) implementation of the MuZero algorithm applied to a vectorized Breakout environment. It is not a typical web or mobile app but a computational pipeline combining environment simulation, search (MCTS), neural networks, and training orchestration.
File structure
• Top-level files:
• README.md – high-level overview and how to run
• config.yaml – hyperparameter and training configuration
• train_torch.py – main training script / controller
• replay_buffer.py – experience storage and sampling
• utils.py – miscellaneous helper functions
• Directory “environment/”: custom vectorized Breakout implementation (parallel environment wrapper)
• Directory “src/”: core algorithmic modules
• mcts.py – Monte Carlo Tree Search in latent space
• networks.py – MuZero networks (representation, dynamics, prediction)
• Directory “static/”: output artifacts (GIF, plots) – useful for visualization but not part of the runtime
README for additional insights
• Running: “python3 train_torch.py”, logs piped to TensorBoard
• Describes deviations from MuZero paper (action mask removal, target network, replay buffer size)
• References to TensorBoard, PyTorch, vectorized environments