facebookresearch/reagent

ReAgent is an archived open-source platform built by Facebook for applying reinforcement learning to real-world problems at scale. Reinforcement learning is a type of AI training where a system learns by taking actions and receiving feedback, rather than being given labeled examples. ReAgent was built to make this approach practical for large-scale recommendation and optimization tasks, such as deciding what to show users in a feed or how to allocate resources. The platform is no longer actively maintained. Facebook's team has moved to a successor project called Pearl, and the README directs users there for ongoing support. ReAgent is preserved as a reference and for teams that built on top of it. When it was active, ReAgent provided a full pipeline: taking raw data, transforming it into training-ready inputs, training AI models using a variety of reinforcement learning algorithms, and then serving those models in production. It was designed for situations where you cannot run a live simulation, so it trained on batches of previously collected data rather than interacting with an environment in real time. It also included tools for evaluating a new policy using old data, which is important when you cannot test a policy by actually deploying it. The platform supported a wide range of algorithms, including several variants of Q-learning for environments with discrete or continuous actions, policy gradient methods, and contextual bandit approaches for simpler decision problems where only a single decision is made rather than a sequence. ReAgent was written in Python and used PyTorch for model training. It supported distributed training for handling large datasets. Installation was possible via Docker or manually, with detailed instructions available in the repository's documentation folder.