Pearl

Why we included this project

Building a reinforcement learning agent that survives contact with a real product is a different job than training one on a clean simulator, and that is the gap Pearl targets. Meta's Applied RL team built it as a modular library where you pick and combine policy learners, replay buffers, exploration strategies, and safety layers, and it supports offline RL, contextual bandits, dynamic action spaces, and history summarization without forcing a rewrite. It is built for production constraints like limited observability, sparse feedback, and high stochasticity, the messy conditions most textbook examples skip. The tutorials cover recommender systems and contextual bandit setups, and the codebase points to real deployments in auction bidding and creative selection, which gives practitioners a sense of where the library actually fits. If you know RL and want a toolkit you can embed into an existing pipeline rather than a research sandbox, Pearl is worth a look.