update
This commit is contained in:
@@ -1,35 +1,123 @@
|
||||
## 思路
|
||||
# DQN
|
||||
|
||||
见[我的博客](https://blog.csdn.net/JohnJim0/article/details/109557173)
|
||||
## 环境
|
||||
## 原理简介
|
||||
DQN是Q-leanning算法的优化和延伸,Q-leaning中使用有限的Q表存储值的信息,而DQN中则用神经网络替代Q表存储信息,这样更适用于高维的情况,相关知识基础可参考[datawhale李宏毅笔记-Q学习](https://datawhalechina.github.io/leedeeprl-notes/#/chapter6/chapter6)。
|
||||
|
||||
python 3.7.9
|
||||
论文方面主要可以参考两篇,一篇就是2013年谷歌DeepMind团队的[Playing Atari with Deep Reinforcement Learning](https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf),一篇是也是他们团队后来在Nature杂志上发表的[Human-level control through deep reinforcement learning](https://web.stanford.edu/class/psych209/Readings/MnihEtAlHassibis15NatureControlDeepRL.pdf)。后者在算法层面增加target q-net,也可以叫做Nature DQN。
|
||||
|
||||
pytorch 1.6.0
|
||||
Nature DQN使用了两个Q网络,一个当前Q网络𝑄用来选择动作,更新模型参数,另一个目标Q网络𝑄′用于计算目标Q值。目标Q网络的网络参数不需要迭代更新,而是每隔一段时间从当前Q网络𝑄复制过来,即延时更新,这样可以减少目标Q值和当前的Q值相关性。
|
||||
|
||||
tensorboard 2.3.0
|
||||
要注意的是,两个Q网络的结构是一模一样的。这样才可以复制网络参数。Nature DQN和[Playing Atari with Deep Reinforcement Learning](https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf)相比,除了用一个新的相同结构的目标Q网络来计算目标Q值以外,其余部分基本是完全相同的。细节也可参考[强化学习(九)Deep Q-Learning进阶之Nature DQN](https://www.cnblogs.com/pinard/p/9756075.html)。
|
||||
|
||||
torchvision 0.7.0
|
||||
https://blog.csdn.net/JohnJim0/article/details/109557173)
|
||||
|
||||
## 使用
|
||||
## 伪代码
|
||||
|
||||
train:
|
||||
<img src="assets/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L0pvaG5KaW0w,size_16,color_FFFFFF,t_70.png" alt="img" style="zoom:50%;" />
|
||||
|
||||
## 代码实战
|
||||
|
||||
### RL接口
|
||||
|
||||
首先是强化学习训练的基本接口,即通用的训练模式:
|
||||
```python
|
||||
python main.py
|
||||
for i_episode in range(MAX_EPISODES):
|
||||
state = env.reset() # reset环境状态
|
||||
for i_step in range(MAX_STEPS):
|
||||
action = agent.choose_action(state) # 根据当前环境state选择action
|
||||
next_state, reward, done, _ = env.step(action) # 更新环境参数
|
||||
agent.memory.push(state, action, reward, next_state, done) # 将state等这些transition存入memory
|
||||
agent.update() # 每步更新网络
|
||||
state = next_state # 跳转到下一个状态
|
||||
if done:
|
||||
break
|
||||
```
|
||||
如上,首先需要循环多个episode训练,在每个episode中,首先需要重置环境,然后开始探索,每个episode加一个MAX_STEPS(也可以使用while not done, 加这个max_steps有时是因为比如gym环境训练目标就是在200个step下达到200的reward),接下来的流程如下:
|
||||
1. agent选择动作
|
||||
2. 环境根据agent的动作反馈出新的state和reward
|
||||
3. agent进行更新,如有memory就会将transition(包含state,reward,action等)存入memory中
|
||||
4. 跳转到下一个状态
|
||||
如果提前done了,就跳出for循环,进行下一个episode的训练。
|
||||
|
||||
eval:
|
||||
|
||||
### 两个Q网络
|
||||
前面讲了Nature DQN中有两个Q网络,一个是policy_net,一个是延时更新的target_net,两个网络的结构是一模一样的,如下(见```model.py```):
|
||||
```python
|
||||
python main.py --train 0
|
||||
```
|
||||
可视化:
|
||||
```python
|
||||
tensorboard --logdir logs
|
||||
```
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
## Torch知识
|
||||
class FCN(nn.Module):
|
||||
def __init__(self, n_states=4, n_actions=18):
|
||||
""" 初始化q网络,为全连接网络
|
||||
n_states: 输入的feature即环境的state数目
|
||||
n_actions: 输出的action总个数
|
||||
"""
|
||||
super(FCN, self).__init__()
|
||||
self.fc1 = nn.Linear(n_states, 128) # 输入层
|
||||
self.fc2 = nn.Linear(128, 128) # 隐藏层
|
||||
self.fc3 = nn.Linear(128, n_actions) # 输出层
|
||||
|
||||
def forward(self, x):
|
||||
# 各层对应的激活函数
|
||||
x = F.relu(self.fc1(x))
|
||||
x = F.relu(self.fc2(x))
|
||||
return self.fc3(x)
|
||||
```
|
||||
输入为state,输出为action,注意根据state和action的维度调整隐藏层的层数,这里设为128
|
||||
|
||||
在```agent.py```中我们定义强化学习算法,包括```choose_action```和```update```两个主要函数,初始化中:
|
||||
```python
|
||||
self.policy_net = FCN(n_states, n_actions).to(self.device)
|
||||
self.target_net = FCN(n_states, n_actions).to(self.device)
|
||||
# target_net的初始模型参数完全复制policy_net
|
||||
self.target_net.load_state_dict(self.policy_net.state_dict())
|
||||
self.target_net.eval() # 不启用 BatchNormalization 和 Dropout
|
||||
# 可查parameters()与state_dict()的区别,前者require_grad=True
|
||||
```
|
||||
可以看到policy_net跟target_net结构和初始参数一样,但在更新的时候target是每隔一段episode更新的,如下(见```main.py```):
|
||||
```python
|
||||
# 更新target network,复制DQN中的所有weights and biases
|
||||
if i_episode % cfg.target_update == 0:
|
||||
agent.target_net.load_state_dict(agent.policy_net.state_dict())
|
||||
```
|
||||
可以调整```cfg.target_update```,注意该变量不要调得太大,否则会收敛很慢,我们最后保存的模型也是这个target_net,如下(见```agent.py```):
|
||||
```python
|
||||
def save_model(self,path):
|
||||
torch.save(self.target_net.state_dict(), path)
|
||||
```
|
||||
### Replay Memory
|
||||
然后就是Replay Memory了,如下(见```memory.py```):
|
||||
```python
|
||||
import random
|
||||
import numpy as np
|
||||
|
||||
class ReplayBuffer:
|
||||
|
||||
def __init__(self, capacity):
|
||||
self.capacity = capacity
|
||||
self.buffer = []
|
||||
self.position = 0
|
||||
|
||||
def push(self, state, action, reward, next_state, done):
|
||||
if len(self.buffer) < self.capacity:
|
||||
self.buffer.append(None)
|
||||
self.buffer[self.position] = (state, action, reward, next_state, done)
|
||||
self.position = (self.position + 1) % self.capacity
|
||||
|
||||
def sample(self, batch_size):
|
||||
batch = random.sample(self.buffer, batch_size)
|
||||
state, action, reward, next_state, done = zip(*batch)
|
||||
return state, action, reward, next_state, done
|
||||
|
||||
def __len__(self):
|
||||
return len(self.buffer)
|
||||
```
|
||||
其实比较简单,主要包括push和sample两个步骤,push是将transitions放到memory中,sample是从memory随机抽取一些transition。
|
||||
|
||||
最后结果如下:
|
||||
|
||||

|
||||
|
||||
## 参考
|
||||
|
||||
[with torch.no_grad()](https://www.jianshu.com/p/1cea017f5d11)
|
||||
|
||||
|
||||
BIN
codes/DQN/assets/rewards_curve_train.png
Normal file
BIN
codes/DQN/assets/rewards_curve_train.png
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 58 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 325 KiB |
Reference in New Issue
Block a user