OpenAI Baselines is OpenAI’s collection of high-quality implementations of reinforcement learning algorithms, created so the research community can replicate results, refine ideas and compare new methods against trustworthy reference implementations. Its DQN implementation and variants score roughly on par with the results in the original published papers. The repository has accumulated more than 16,000 stars on GitHub. OpenAI maintains it in maintenance mode, meaning it still receives bug fixes and minor updates, with the last code push in 2024.
Baselines covers the classic deep reinforcement learning algorithms that defined the field, each implemented to be readable, reusable and close to its paper: DQN and its variants for value-based control, PPO and TRPO for policy optimization, A2C and ACKTR for actor-critic methods, DDPG for continuous control, and additions such as HER (Hindsight Experience Replay) and GAIL for imitation learning. Because every algorithm shares common infrastructure for environments, logging and evaluation, researchers can swap algorithms on the same task and trust that differences come from the method, not from incidental implementation details. For years it served as the standard starting point for RL experiments and coursework.
OpenAI Baselines is completely free and open source under the MIT license. There are no paid plans or hosted versions.
Baselines is for reinforcement learning researchers, students and engineers who need dependable reference implementations of classic RL algorithms to learn from, benchmark against or build new research on.