Pearl

Active
GitHub Jupyter Notebook MIT

Description

A production-ready Reinforcement Learning AI Agent Library from Meta with comprehensive algorithm implementations.

Key Features

  • Modular design for flexible agent construction via mix-and-match components
  • Dynamic action spaces and offline learning capabilities
  • Intelligent neural exploration for sparse feedback environments
  • Safe decision-making with safety-constrained actor-critic algorithms
  • History summarization and data augmentation for production robustness
  • Component serialization via PyTorch state dicts for model persistence

Use Cases

💡 Building recommendation systems that learn from user interaction over time
💡 Auction bidding optimization with safety constraints
💡 Creative selection and A/B testing with RL agents
💡 Research on contextual bandits and sequential decision making
💡 Deploying RL agents in environments with limited observability

Tags

reinforcement-learning meta production algorithm python

Categories

🤖 Agent Framework

Quick Start

1. Clone the repo and install: git clone https://github.com/facebookresearch/Pearl.git && cd Pearl && pip install -e . 2. Create a PearlAgent with a policy learner (e.g., DeepQLearning) and replay buffer. 3. Connect to an environment, call agent.reset(), then loop: agent.act(), env.step(), agent.observe(), agent.learn(). See tutorials for recommender systems, contextual bandits, and Frozen Lake examples.
Visit GitHub

Related Projects