Skip to main content

Ctrl+K

BreakRL

中文版

最小演示
多臂老虎机：探索与利用
马尔可夫决策过程：序列决策形式化
时序差分学习：长期价值估计
DQN：神经网络价值学习
Policy Gradient / REINFORCE：直接优化策略
Actor-Critic / A2C：价值辅助策略更新
PPO：约束策略更新
SAC：最大熵连续控制
离线强化学习：CQL 与 IQL
模型式强化学习：从环境模型到 Dyna-Q
Decision Transformer：序列建模的强化学习
RLHF：从偏好到奖励
DPO：不训奖励模型的偏好优化
GRPO 与 RLVR：可验证奖励
离线强化学习正文（PDF）
RL 失败模式图鉴

English

Home
Minimum demo
Multi-Armed Bandits: Exploration vs Exploitation
Markov Decision Processes
Temporal-Difference Learning
DQN: Neural Value Learning
Policy Gradient / REINFORCE
Actor-Critic / A2C
PPO: Constrained Policy Updates
SAC: Maximum-Entropy Continuous Control
Offline RL: CQL and IQL
Model-Based RL: Dyna-Q
Decision Transformer: RL as Sequence Modeling
RLHF: From Preferences to Rewards
DPO: Preference Optimization without a Reward Model
GRPO and RLVR: Verifiable Rewards
Offline RL chapter (PDF)
Failure Atlas

Index

By BreakRL Contributors

© Copyright 2026, BreakRL Contributors.

BreakRL 是一份双语、失败优先的强化学习讲义。正文 CC BY 4.0，代码 MIT。

BreakRL is a bilingual, failure-first set of RL lecture notes. Text CC BY 4.0. Code MIT.