Keras-RL provides a collection of state-of-the-art deep reinforcement learning algorithms implemented in Python, designed for seamless integration with the Keras deep learning library. It targets researchers and developers looking to experiment with and extend RL algorithms, offering easy integration with OpenAI Gym environments for rapid prototyping and evaluation.

How It Works

The library implements various RL algorithms by extending abstract classes, allowing users to easily define custom environments and agents. It leverages Keras for neural network construction and training, enabling the use of built-in or custom Keras callbacks and metrics for monitoring and controlling the learning process.

Quick Start & Requirements

• Install via pip: pip install keras-rl
• Additional dependencies for examples: gym, h5py, Pillow, gym[atari]
• Example usage: python examples/dqn_cartpole.py
• Documentation: keras-rl documentation

Highlighted Details

• Implements DQN, Double DQN, DDPG, Continuous DQN (NAF), CEM, Dueling DQN, and Deep SARSA.
• Supports integration with Weights & Biases (wandb) for training visualization via WandbLogger callback.
• Allows easy extension for custom environments and algorithms.
• Sample weights are available for download.

Maintenance & Community

• The project is hosted on GitHub.
• Users can file issues or submit pull requests for contributions.

Licensing & Compatibility

• The library is released under the MIT License.
• Compatible with Keras and OpenAI Gym.

Limitations & Caveats

• Asynchronous Advantage Actor-Critic (A3C) and Proximal Policy Optimization (PPO) algorithms are listed as not yet implemented.
• The project has not been updated since 2017, indicating potential compatibility issues with newer Keras or TensorFlow versions.