引用次数 394
速度、稳定
共同架构：DQN
| 方案 | 原理 | 解决问题 |
|---|---|---|
| Double DQN | decouple selection and evaluation of actions | overestimation bias |
| Prioritized DDQN | Prioritized Experience Replay | data efficiency |
| Dueling Prioritized DDQN | Dueling network arch | generalize across actions |
| A3C | Multi-step，使用截断的 n 步折扣奖励 | 更快，shift bias-variance trade-off, propagate rewards faster |
| Distributional DQN | learn distribution of returns | 随机性? |
| Noisy DQN | stochastic network layer | exploration |
Q-learning
$$ \Delta = R_{t+1} + \gamma_{t+1} \max\limits_{a'} q_{\bar{\theta}} (S_{t+1}, a') - q_{\theta} (S_t, A_T) $$
Double Q-learning
$$ \Delta = R_{t+1} + \gamma_{t+1} q_{\bar{\theta}} (S_{t+1}, \argmax\limits_{a'} q_{\theta} (S_{t+1}, a')) - q_{\theta} (S_t, A_T) $$
Prioritized Replay
$$ p_t \propto |\Delta|^\omega $$
Deuling network https://arxiv.org/abs/1511.06581.pdf
Page Not Found
Try to search through the entire repo.