Ray RLlib: Scalable Reinforcement Learning

https://github.com/ray-project/ray

A high-performance distributed execution engine

Ray is a flexible, high-performance distributed execution framework.

Ray comes with libraries that accelerate deep learning and reinforcement learning development:

• Ray Tune: Hyperparameter Optimization Framework
• Ray RLlib: Scalable Reinforcement Learning

More Information

• Documentation
• Tutorial
• Blog
• Ray paper
• Ray HotOS paper

Ray RLlib: Scalable Reinforcement Learning

Ray RLlib is an RL execution toolkit built on the Ray distributed execution framework. See the user documentation and paper.

RLlib includes the following reference algorithms:

• Proximal Policy Optimization (PPO) which is a proximal variant of TRPO.
• Policy Gradients (PG).
• Asynchronous Advantage Actor-Critic (A3C).
• Deep Q Networks (DQN).
• Deep Deterministic Policy Gradients (DDPG, DDPG2).
• Ape-X Distributed Prioritized Experience Replay, including both DQN and DDPG variants.
• Evolution Strategies (ES), as described in this paper.

These algorithms can be run on any OpenAI Gym MDP, including custom ones written and registered by the user.