google/dopamine
Dopamine
Dopamine is a reinforcement learning research framework designed for prototyping and testing algorithms across diverse simulated environments. It provides an agent development toolkit that utilizes a flat class hierarchy to facilitate the creation and extension of learning agents.
The framework includes a standardization layer via environment wrappers that connect agents to various physics simulations and gaming environments. It also features a high-performance experience replay buffer for storing and sampling transition data to improve training stability, alongside a dedicated hyperparameter management system that uses external files to ensure reproducible research.
The library covers broader capabilities for experiment management, including metric-based logging for training statistics, checkpoint-based state recovery for model weights, and performance benchmarking. It also provides utilities for analyzing training progress and visualizing agent behavior through images and video.
Features
- Reinforcement Learning Research Frameworks - Acts as a research framework for prototyping and testing RL algorithms across diverse simulated environments.
- Experience Replay Buffers - Implements a high-performance replay buffer to store and sample transition data for training stability.
- Hyperparameter Configurations - Provides tools for managing and tuning model hyperparameters via external files for reproducible research.
- Environment Wrappers - Implements environment wrappers to standardize inputs and outputs across diverse simulation interfaces.
- RL Training Workflows - Provides standardized processes for training reinforcement learning agents within simulated environments.
- RL Agent Implementation Frameworks - Ships frameworks to implement agents for both discrete and continuous control tasks.
- RL Experiment Configurations - Provides declarative setups for managing hyperparameters and configurations to ensure research reproducibility.
- Reinforcement Learning Frameworks - Provides a complete framework and toolkit for prototyping reinforcement learning agents in decision-making tasks.
- Simulation Environments - Links learning agents to diverse simulation interfaces using wrappers to standardize data exchange.
- Agent Development - Provides a specialized toolkit and class hierarchy for developing and extending custom reinforcement learning agents.
- Agent Development - Provides platforms and interfaces to build and customize reinforcement learning agents through subclassing.
- Experiment Logging - Records training statistics and saves model checkpoints to local directories for analysis and state recovery.
- Training Progress Monitoring - Tracks and visualizes loss metrics and agent statistics to monitor reinforcement learning progress.
- Model Checkpoints - Supports importing weights from agent checkpoints to resume training or evaluate performance.
- Agent Architectures - Supports building new RL algorithms using a flat class hierarchy for extensible agent architecture.
- Performance Benchmarking - Provides mechanisms to measure RL algorithm accuracy using standardized environment benchmarks.
- Training Checkpointers - Provides mechanisms for saving and resuming model weights and agent states to ensure training continuity.
- Training Log Analysis - Includes utilities for loading and visualizing training logs to monitor agent progress and loss curves.
- External Configuration Loading - Allows loading hyperparameters from external files and command-line arguments to ensure experiment reproducibility.
- Agent - Utilizes a flat class hierarchy to make the addition of new RL algorithms straightforward.
- Agent Performance Visualizers - Generates videos and images of agents interacting with environments to analyze behavioral reliability.
- Training Metric Pipelines - Pipes in-iteration losses and performance statistics to external visualization tools and local directories.
- Machine-Learning-Frameworks - Framework for rapid prototyping of reinforcement learning algorithms.
- Perception and Machine Learning - Research framework for reinforcement learning prototyping.
- Reinforcement Learning - Research framework for prototyping reinforcement learning algorithms.
