GitHub - pfnet/pfrl: PFRL: a PyTorch-based deep reinforcement learning library

reinforcement-learningdeep-learningpytorchlibraryopen-source

Abstraction: PyTorch open-source library implementing state-of-the-art deep RL algorithms

Key points:

PFRL (Preferred Networks Reinforcement Learning) is a PyTorch-based library implementing 11 major deep RL algorithms: DQN, Categorical DQN, Rainbow, IQN, DDPG, A3C/A2C, ACER, PPO, TRPO, TD3, and SAC.
Supports discrete and continuous action spaces, recurrent models, batch training, and CPU async training depending on algorithm.
Provides pretrained model zoos for Atari (DQN, IQN, Rainbow, A3C) and MuJoCo (DDPG, TRPO, PPO, TD3, SAC) environments.
Also implements auxiliary techniques: NoisyNet, Prioritized Experience Replay, Dueling Networks, Normalized Advantage Function, and Deep Recurrent Q-Network.
Based on ChainerRL (cited JMLR 2021 paper by Fujita et al.); compatible with any environment implementing OpenAI Gym's reset/step interface.

Connections: Pfrl · Preferred Networks · Pytorch · Reinforcement Learning · Deep Reinforcement Learning