vwxyzjn/cleanrl

Features

• PyTorch Training Frameworks - Provides a PyTorch-based framework for training agents in both continuous and discrete action spaces.
• Reinforcement Learning Research Frameworks - Provides a comprehensive framework for prototyping, testing, and benchmarking online reinforcement learning algorithms.
• RL Agent Implementation Frameworks - Provides a framework for creating standalone, single-file implementations of RL algorithms to simplify prototyping and research.
• Soft Actor-Critic Implementations - Implements a stochastic policy that maximizes entropy to ensure stable learning in continuous spaces.
• Benchmarking Suites - Provides a set of reproducible algorithm implementations to evaluate agent performance on standard benchmarks.
• Continuous Control Actors - Implements deterministic actor networks for executing continuous actions in high-dimensional physics simulations.
• Continuous Control Training - Develops agents for high-dimensional continuous action spaces using algorithms like SAC and TD3.
• Parallel Q-Learning Implementations - Provides a Q-learning implementation using multiple parallel environments to improve training efficiency.
• Twin Delayed DDPG Implementations - Provides a continuous control implementation using clipped double Q-learning and delayed policy updates.
• Observation Processors - Provides observation processors that handle diverse input types, including image frames and state vectors, for environment compatibility.
• Visual Observation Pipelines - Implements a pipeline for handling raw Atari pixel observations using convolutional layers and standard image pre-processing.
• Pixel-Based RL Training - Processes raw image observations with convolutional layers and recurrent networks to train agents in visual environments.
• Deep Deterministic - Provides a deterministic policy implementation for continuous action spaces using an actor-critic architecture.
• Reinforcement Learning - Offers a collection of standalone, reproducible implementations of online reinforcement learning algorithms.
• Environment Wrappers - Implements toolsets for interfacing with simulation environments to preprocess observations and actions for training.
• Vectorized Environments - Runs multiple simulation instances in parallel via vectorized environment wrapping to increase data collection throughput.
• PPO Implementations - Provides a stable training implementation of Proximal Policy Optimization using a clipped surrogate objective.
• Categorical DQN Implementations - Implements a distributional RL agent that learns a probability distribution over return values.
• Rainbow DQN Implementations - Implements a comprehensive DQN agent combining prioritized replay and distributional learning enhancements.
• Single-File Algorithm Implementations - Encapsulates reinforcement learning algorithms in standalone scripts to eliminate complex class hierarchies and dependencies.
• Clipped Double Q-Learning - Implements clipped double Q-learning to mitigate overestimation bias during continuous action space training.
• Distributed Training - Provides mechanisms to distribute training data across multiple GPUs and vectorized environments to reduce training time.
• Experiment Logging - Implements a system for logging training statistics and model checkpoints to external dashboards.
• Intrinsic Reward Modules - Uses prediction errors between target and predictor networks to generate intrinsic rewards for exploring sparse environments.
• Hardware-Accelerated Simulations - Provides an interface for GPU-accelerated physics engines allowing thousands of simultaneous simulation environments.
• Agent Performance Evaluators - Evaluates agent learning by recording gameplay videos and logging training metrics to dashboards.
• Phasic Policy Gradient Implementations - Implements a training approach that alternates between policy and value function phases using a replay buffer for sample efficiency.
• Hyperparameter Tuning - Implements automated sampling and pruning workflows to find optimal algorithm configurations across multiple environments.
• Hyperparameter Tuning - Ships a system to identify optimal hyperparameters that generalize across multiple tasks using custom samplers and pruners.
• Self-Play Architectures - Provides a vectorized architecture that enables agents to compete or cooperate through multi-agent self-play.
• Episodic Memory Modules - Employs transformer architectures to enable agents to remember and leverage past observations in partially observable environments.
• Temporal State Memory - Uses memory layers to process sequential observations in pixel-based environments, eliminating the need for stacked frames.
• Recurrent Neural Networks - Integrates recurrent neural network layers within policy gradients to handle sequential dependencies and partially observable environments.
• Sampling Accelerators - Uses high-performance environment wrappers to increase data collection throughput during agent training.
• Policy Clipping - Implements a clipped surrogate objective to ensure stable policy updates in proximal policy optimization.
• Cloud RL Experimentation - Scales reinforcement learning training by deploying containers to cloud batch services for parallel execution across seeds.
• Cloud Batch Processing - Executes thousands of training trials simultaneously by packaging code into containers for cloud batch services.
• Parallel Execution - Schedules multiple independent training trials across CPU cores or cluster nodes using separate process instances.
• Cloud Infrastructure Automation - Automates the creation and removal of remote computing resources using infrastructure-as-code tools.
• Containerized Deployments - Packages environments and dependencies into container images for scalable execution across GPU and CPU clusters.
• Remote Compute Job Submission - Deploys training jobs to remote compute instances utilizing specific CPU or GPU hardware profiles.
• Experiment Orchestrators - Ships a system for scheduling and executing large-scale training trials across cloud clusters with containerized workloads.
• Automated Experiment Runners - Provides a tool for scheduling and executing multiple training trials across various environments and seeds.
• Real-Time Monitoring Dashboards - Provides real-time monitoring dashboards to track agent performance metrics during the training process.
• Reinforcement Learning - Single-file implementations of deep reinforcement learning algorithms.

Star 历史