This project provides a Python framework for reinforcement learning (RL) experimentation, offering implementations of state-of-the-art algorithms and easy integration of new environments. It is targeted at researchers and developers looking to quickly prototype and benchmark RL algorithms.

How It Works

Coach features a modular design, decoupling core RL components like algorithms, environments, network architectures, and exploration policies. This allows for straightforward extension and reuse of existing components. The framework utilizes a "preset" system for defining and running experiments, simplifying the process of training agents and reproducing results.

Quick Start & Requirements

• Installation: Recommended within a virtual environment (virtualenv -p python3 coach_env). Install via pip: pip3 install rl_coach. For development, clone the repo and install with pip3 install -e ..
• Prerequisites: Requires Ubuntu 16.04 LTS and Python 3.5. Extensive system dependencies include cmake, python3-tk, python-opencv, Boost libraries, SciPy dependencies (libblas-dev, liblapack-dev, libatlas-base-dev, gfortran), PyGame dependencies, Dashboard dependencies, and Gym dependencies.
• GPU: Installs tensorflow-gpu by default if a GPU is present; otherwise, installs an Intel-Optimized TensorFlow.
• Documentation: Tutorials and framework documentation are available.

Highlighted Details

• Supports a wide range of RL algorithms, including DQN, A3C, DDPG, PPO, SAC, and TD3.
• Offers implementations for various memory types (HER, PER) and exploration techniques.
• Includes a dashboard for visualizing training signals and debugging.
• Supports distributed multi-node training and batch reinforcement learning from datasets.

Maintenance & Community

• DISCONTINUATION OF PROJECT: Intel has ceased maintenance, bug fixes, and new releases. The project is no longer actively supported by Intel. Users are encouraged to fork the project for independent development or maintenance.

Licensing & Compatibility

• The project is released as reference code for research purposes. No specific license is mentioned in the README, but it is stated to be "not an official Intel product."

Limitations & Caveats

• The project is discontinued and will no longer be maintained by Intel.
• Installation is primarily tested on Ubuntu 16.04 LTS with Python 3.5, and requires significant system-level dependencies.
• MuJoCo version compatibility issues exist between OpenAI Gym (v1.5) and Robosuite (v2.0), potentially requiring separate virtual environments.