NeatRL
Readable single-file PyTorch implementations of deep RL algorithms
What is NeatRL?
Readable single-file PyTorch implementations of deep RL algorithms
Overview of NeatRL
What it is
NeatRL is a Python library for reinforcement learning algorithms. It provides readable single-file implementations of deep RL algorithms using PyTorch. The library supports Gymnasium environments and Atari games, and includes Weights & Biases integration for experiment tracking. It is designed for research and experimentation.
Key features
- Single-file implementations: each algorithm is contained in one self-contained file.
- Experiment tracking: built-in support for Weights & Biases logging.
- Gymnasium compatible: works with Gymnasium environments.
- Atari support: full support for Atari games with automatic CNN architectures.
- Parallel training: vectorized environments for faster data collection.
- Easy to extend: modular design for adding new algorithms.
- Video recording: automatic video capture and WandB integration.
Use cases
- Research and experimentation with reinforcement learning algorithms.
- Training RL agents on discrete action space environments such as CartPole and FrozenLake.
- Training RL agents on continuous control tasks such as Pendulum and HalfCheetah.
- Training RL agents on Atari games with automatic preprocessing.
- Comparing different algorithms (DQN, PPO, SAC, TD3, DDPG, A2C, etc.) on standard benchmarks.
Getting started
Install NeatRL via pip: bash pip install neatrl pip install 'neatrl[cli]' # Live terminal dashboard (rich, psutil) Optional environment extras: bash pip install 'neatrl[atari]' # Atari games (ALE) pip install 'neatrl[box2d]' # Box2D environments (LunarLander, BipedalWalker, …) pip install 'neatrl[classic]' # Classic control (CartPole, Pendulum, …) pip install 'neatrl[mujoco]' # MuJoCo continuous control Quick start example training DQN on CartPole: python from neatrl.dqn_mlp import train_dqn model = train_dqn(env_id='CartPole-v1', total_timesteps=10000, seed=42)
License and pricing
The code is licensed under Apache License 2.0. Documentation, diagrams, benchmark results and other non-code assets are licensed under CC BY 4.0. The project is open source.
Based on the project's public README and repository metadata.