# RL-Study 强化学习算法实现与学习笔记,基于赵世钰老师《Mathematical Foundations of Reinforcement Learning》。 ## 项目结构 ``` RL-Study/ ├── Lecture slides/ # 课程幻灯片 │ ├── slidesForMyLectureVideos/ # 配套视频课件 │ └── slidesContinuouslyUpdated/ # 持续更新的课件 ├── Notebooks/ # Jupyter 学习笔记 │ ├── C1.ipynb ~ C10.ipynb # 各章节推导与实验 │ ├── SAC.ipynb # SAC (Soft Actor-Critic) 算法 │ └── *_training_results.png # 训练结果可视化 ├── RawBook/ # 原书资源 ├── RL_Algothrithms/ # 核心算法实现 │ ├── agents/ # 智能体实现 │ │ ├── a2c.py # A2C (Advantage Actor-Critic) │ │ ├── qac.py # QAC (Q-Value Actor-Critic) │ │ ├── off_pac.py # Off-PAC (Off-Policy Actor-Critic) │ │ ├── ddpg.py # DDPG (Deep Deterministic Policy Gradient) │ │ └── dpac.py # DPAC (Deterministic Policy Actor-Critic) │ ├── networks.py # 离散动作空间网络 (Actor, QCritic, VCritic) │ ├── networks_cont.py # 连续动作空间网络 (ContActor, ContQCritic) │ ├── utils.py # 工具函数 │ ├── main.py # 离散动作空间训练入口 (CartPole-v1) │ ├── disp_main.py # 离散动作空间多算法对比 │ └── cont_main.py # 连续动作空间训练入口 (Pendulum-v1) └── TRPO/ # TRPO (Trust Region Policy Optimization) 独立实现 ├── models.py # ActorNet (高斯策略), CriticNet (值函数) ├── utils.py # RolloutBuffer, GAE, 共轭梯度, FVP ├── agent.py # TRPO 智能体 └── main.py # TRPO 训练入口 (Pendulum-v1) ``` ## 已实现算法 ### 离散动作空间(CartPole-v1) | 算法 | 文件 | 说明 | | ------- | ------------------------------------------------------ | ------------------------------------------------------ | | A2C | [agents/a2c.py](RL_Algothrithms/agents/a2c.py) | Advantage Actor-Critic,同步版本,V-critic,带熵正则化 | | QAC | [agents/qac.py](RL_Algothrithms/agents/qac.py) | Q-Value Actor-Critic,On-policy SARSA 风格,支持 GPU | | Off-PAC | [agents/off_pac.py](RL_Algothrithms/agents/off_pac.py) | Off-Policy Actor-Critic,带重要性采样,epsilon 探索 | ### 连续动作空间(Pendulum-v1) | 算法 | 文件 | 说明 | | ---- | ------------------------------------------------ | -------------------------------------------------------------- | | DDPG | [agents/ddpg.py](RL_Algothrithms/agents/ddpg.py) | Deep Deterministic Policy Gradient,离策略,带目标网络和软更新 | | DPAC | [agents/dpac.py](RL_Algothrithms/agents/dpac.py) | Deterministic Policy Actor-Critic,在策略 | | TRPO | [agent.py](TRPO/agent.py) | Trust Region Policy Optimization,共轭梯度法 + 线搜索 + GAE | ## 环境配置 ```bash pip install torch numpy matplotlib gymnasium ``` ## 快速开始 ### 离散动作空间(CartPole-v1) ```bash cd RL_Algothrithms python main.py --agent a2c # 训练 A2C python main.py --agent qac # 训练 QAC python disp_main.py # 多算法对比训练 ``` ### 连续动作空间(Pendulum-v1) ```bash cd RL_Algothrithms python cont_main.py # 训练 DDPG 和 DPAC ``` ### TRPO ```bash cd TRPO python main.py # 训练 TRPO ``` ## 关于原书 - **书名**: Mathematical Foundations of Reinforcement Learning - **作者**: Shiyu Zhao (Westlake University) - **GitHub**: [MathFoundationRL/Book-Mathematical-Foundation-of-Reinforcement-Learning](https://github.com/MathFoundationRL/Book-Mathematical-Foundation-of-Reinforcement-Learning) - **B站**: [赵世钰老师频道](https://space.bilibili.com/2044042934) - **YouTube**: [课程列表](https://youtube.com/playlist?list=PLEhdbSEZZbDaFWPX4gehhwB9vJZJ1DNm8) ## License MIT License(代码部分)