abcdRL (Implement a RL algorithm in four simple steps)

abcdRL is a Modular Single-file Reinforcement Learning Algorithms Library that provides modular design without strict and clean single-file implementation.

Adam

Understand the full implementation details of the algorithm in a single file quickly when reading the code; Benefit from a lightweight modular design, only need to focus on a small number of modules when modifying the algorithm.

Ref

abcdRL mainly references the single-file design philosophy of vwxyzjn/cleanrl and the module design of PaddlePaddle/PARL.

Roadmap🗺️ #57

🗽 Design Philosophy

• "Copy📋", not "Inheritance🧬"
• "Single-file📜", not "Multi-file📚"
• "Features reuse🛠", not "Algorithms reuse🖨"
• "Unified logic🤖", not "Unified interface🔌"

✅ Implemented Algorithms

Weights & Biases Benchmark Report ➡️ report.abcdrl.xyz

• Deep Q Network (DQN) _{dqn_torch.py, dqn_tf.py, dqn_atari_torch.py, dqn_atari_tf.py}
• Deep Deterministic Policy Gradient (DDPG) _{ddpg_torch.py}
• Twin Delayed Deep Deterministic Policy Gradient (TD3) _{td3_torch.py}
• Soft Actor-Critic (SAC) _{sac_torch.py}
• Proximal Policy Optimization (PPO) _{ppo_torch.py}

---

• Double Deep Q Network (DDQN) _{ddqn_torch.py, ddqn_tf.py}
• Prioritized Deep Q Network (PDQN) _{pdqn_torch.py, pdqn_tf.py}

Citing abcdRL

@misc{zhao_abcdrl_2022,
    author = {Yanxiao, Zhao},
    month = {12},
    title = {{abcdRL: Modular Single-file Reinforcement Learning Algorithms Library}},
    url = {https://github.com/sdpkjc/abcdrl},
    year = {2022}
}

---

Last update: 2023-03-01