Research platform for Human-in-the-loop learning & Multi-Agent Reinforcement Learning

HILL and MARL Training

Our SDK includes base implementations for multiple RL algorithms such as DQN, TD3, A2C and PPO for both single- and multi-agent reinforcement learning (MARL). For human-in-the-loop learning, Cogment Verse includes different paradigms like learning from demonstrations (behavior cloning), learning from human interventions, and learning from explicit human feedback (RLHF). More importantly, the SDK makes it easy to implement any asynchronous centralized training / decentralized execution algorithm.

Interactive Web Application

A built-in interactive web application enables seamless human participation in episodes during the training and validation processes. Additional environments can be made interactive with very little web development required..

Collaboration Patterns

Library of predefined roles for agents or humans facilitating the implementation of multiple collaboration patterns: co-players, teacher/student dual control, evaluators, etc.

Extensive Environment Library

The existing environment library includes the standard OpenAI Gym environments as well as board games, card games, and other multi-agent environments from Petting Zoo. Other environments are gradually being added, such as Isaac Gym or Overcooked-AI. Implementing your own interactive custom environments is also quite straightforward.

Experiments Management

Monitor and manage reproducible experiments, thanks to:

• A powerful configuration management system, including hyperparameters and seeds, built on Hydra.
• Streamlined experiment tracking using mlflow.
• Trained model versioning and storage via the Cogment Model Registry.
• Episode data management from the Cogment Trial Datastore.