This repository provides PyTorch implementations of classic deep reinforcement learning algorithms, targeting researchers and practitioners seeking clear, educational code. It offers implementations of DQN, DDPG, SAC, A2C, PPO, and TRPO, aiming to simplify learning and experimentation in the field.

How It Works

The project implements core RL algorithms using PyTorch, structuring the code into distinct modules for agents, models, utilities, and training scripts. This modular design promotes clarity and maintainability, allowing users to easily understand and modify specific components of each algorithm. The use of PyTorch facilitates GPU acceleration and flexible neural network definition.

Quick Start & Requirements

• Install rl_utils module: pip install -e .
• Prerequisites: PyTorch (>=1.0.1), Gym (>=0.12.5), mujoco-py, opencv-python, cloudpickle.
• Additional installs: gym[atari], gym[box2d].
• Mujoco installation requires following official instructions.
• Training: cd rl_algorithms/<target_algo_folder>/ && python train.py --<arguments>
• Demo: cd rl_algorithms/<target_algo_folder>/ && python demo.py --<arguments>
• Official docs/demos are not explicitly linked, but code structure suggests per-algorithm folders.

Highlighted Details

• Implements DQN (Basic, Double Q, Dueling Network), DDPG, A2C, TRPO, PPO, and SAC.
• PPO supports Atari and Mujoco environments; TRPO is noted as more stable.
• DDPG has been re-implemented with support for more results.
• Future plans include prioritized experience replay and removal of OpenAI baseline preprocessing.

Maintenance & Community

• Last update mentioned is November 2019.
• No community links (Discord/Slack) or roadmap are provided.

Licensing & Compatibility

• The README does not explicitly state a license. Given the dependencies and common practices in RL research, it's likely permissive but requires verification.

Limitations & Caveats

The repository's last update was in late 2019, indicating potential staleness regarding newer RL advancements or PyTorch API changes. The absence of explicit licensing information and community channels may hinder adoption and support.