facebookresearch/ReAgent
ReAgent
ReAgent is a reinforcement learning platform designed for training, deploying, and evaluating reinforcement learning models and contextual bandit systems for large-scale decision making. It provides a comprehensive suite of tools that spans the entire workflow from initial feasibility analysis to production serving.
The system includes a deep reinforcement learning training framework for distributed off-policy algorithms and a specialized model serving layer for high-volume production inference. It distinguishes itself with a counterfactual policy evaluator for estimating performance using historical data and a feature analysis tool to determine if a domain is suitable for batch reinforcement learning.
The platform covers several core capability areas, including behavioral cloning for policy bootstrapping, distributed training workflows with integrated data preprocessing, and domain suitability analysis through state and action importance quantification.
Features
- RL Training Workflows - A comprehensive distributed environment for training off-policy RL algorithms with integrated preprocessing and cloning.
- Behavioral Cloning Toolkits - Implements behavioral cloning to initialize a learning policy by mimicking a logging policy.
- Distributed Training Coordination - Provides mechanisms for synchronizing large-scale deep reinforcement learning training across multiple compute nodes.
- Imitation Learning Bootstrapping - Implements imitation learning to initialize policies using expert datasets, ensuring stability before reinforcement learning begins.
- Off-Policy Corrections - Calculates policy updates using importance sampling ratios for off-policy correction from historical data buffers.
- Optimized Model Serving - Deploys trained reasoning models through a low-latency, high-performance serving layer for real-time decision making.
- Model Serving - Provides the infrastructure for deploying trained reinforcement learning models to production inference endpoints.
- Reinforcement Learning - A platform for training, deploying, and evaluating reinforcement learning models and contextual bandit systems.
- Production Serving Layers - Deploys trained reinforcement learning models to a production serving layer for large scale decision making.
- Model Serving - Provides an optimized infrastructure for deploying trained reasoning models for high-volume production inference.
- Offline Policy Evaluators - Estimates the performance of new policies using logged historical data to avoid risky live deployments.
- Batch Learning Feasibility Analysis - Quantifies the influence of state and action variables to determine if a problem is suitable for batch learning.
- RL Domain Analysis Tools - Provides a framework for assessing state and action importance to determine if a domain is suitable for batch RL.
- RL Domain Suitability Analysis - Evaluates state and action feature importance to determine if a problem is appropriate for batch reinforcement learning.
- Batch RL Suitability Assessments - Analyzes state and action feature importance to determine if a problem is suitable for batch reinforcement learning.
- Machine Learning Frameworks - End-to-end platform for applied reinforcement learning.
- Perception and Machine Learning - End-to-end platform for applied reinforcement learning.
Star history
How this analysis was created: This summary and feature list were written by an AI model that read the project's README and public documentation pages. Each feature links to the documentation it came from; stars, license and language come straight from the GitHub API. The model does not read the source code, and the analysis is refreshed when the project is re-analysed. Learn more on our About page.
