ReAgent is an end-to-end platform for applied reinforcement learning and decision-making systems built in Python using PyTorch and TorchScript.
The platform addresses the challenge of training RL agents in production environments where simulators are unavailable and data collection is expensive. It solves this through offline batch training on historical data combined with off-policy algorithms, allowing teams to develop and evaluate policies without live deployment risk. The tool includes data preprocessing, feature transformation, distributed training infrastructure, counterfactual policy evaluation techniques to estimate policy performance from logged data, and optimized serving via TorchScript.
ReAgent suits large-scale recommendation and optimization tasks where you have access to historical interaction data but cannot easily simulate environments. The platform is particularly valuable for teams building recommender systems or contextual decision systems at scale. It implements a comprehensive set of algorithms spanning classic off-policy deep RL methods like DQN variants and SAC, specialized algorithms for ranking and slate recommendation, contextual bandit approaches, and counterfactual evaluation techniques for policy assessment without deployment.
The project is officially archived and no longer maintained. The development team has directed users to Pearl, a successor library positioned as production-ready reinforcement learning infrastructure by the same organization.
