update
@@ -1,6 +1,7 @@
|
||||
# DQN
|
||||
#TODO
|
||||
|
||||
## 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/easy-rl/#/chapter6/chapter6)。
|
||||
|
||||
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。
|
||||
@@ -15,7 +16,7 @@ https://blog.csdn.net/JohnJim0/article/details/109557173)
|
||||
|
||||
<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
|
||||
## 代码实现
|
||||
|
||||
### RL接口
|
||||
|
||||
@@ -24,23 +25,26 @@ https://blog.csdn.net/JohnJim0/article/details/109557173)
|
||||
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
|
||||
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:
|
||||
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,或者是当完成一个episode的步数较多时也可以设置,基本流程跟所有伪代码一致,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
|
||||
2. 环境根据agent的动作反馈出next_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。
|
||||
5. 如果done了,就跳出循环,进行下一个episode的训练。
|
||||
|
||||
想要实现完整的算法还需要创建Qnet,Replaybuffer等类
|
||||
|
||||
### 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```):
|
||||
|
||||
上文讲了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```),注意DQN使用的Qnet就是全连接网络即FCH:
|
||||
```python
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
@@ -62,30 +66,12 @@ class FCN(nn.Module):
|
||||
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
|
||||
输入为state_dim,输出为action_dim,包含一个128维度的隐藏层,这里根据需要可增加隐藏层维度和数量,然后一般使用relu激活函数,这里跟深度学习的网路设置是一样的。
|
||||
|
||||
### Replay Buffer
|
||||
|
||||
然后就是Replay Memory了,其作用主要是是克服经验数据的相关性(correlated data)和非平稳分布(non-stationary distribution)问题,实现如下(见```memory.py```):
|
||||
|
||||
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(state_dim, action_dim).to(self.device)
|
||||
self.target_net = FCN(state_dim, action_dim).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
|
||||
@@ -111,11 +97,120 @@ class ReplayBuffer:
|
||||
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:
|
||||
参数capacity表示buffer的容量,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。
|
||||
|
||||

|
||||
### Agent类
|
||||
|
||||
在```agent.py```中我们定义强化学习算法类,包括```choose_action```(选择动作,使用e-greedy策略时会多一个```predict```函数,下面会将到)和```update```(更新)等函数。
|
||||
|
||||
在类中建立两个网络,以及optimizer和memory,
|
||||
|
||||
```python
|
||||
self.policy_net = MLP(state_dim, action_dim,hidden_dim=cfg.hidden_dim).to(self.device)
|
||||
self.target_net = MLP(state_dim, action_dim,hidden_dim=cfg.hidden_dim).to(self.device)
|
||||
for target_param, param in zip(self.target_net.parameters(),self.policy_net.parameters()): # copy params from policy net
|
||||
target_param.data.copy_(param.data)
|
||||
self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg.lr)
|
||||
self.memory = ReplayBuffer(cfg.memory_capacity)
|
||||
```
|
||||
然后是选择action:
|
||||
|
||||
```python
|
||||
def choose_action(self, state):
|
||||
'''选择动作
|
||||
'''
|
||||
self.frame_idx += 1
|
||||
if random.random() > self.epsilon(self.frame_idx):
|
||||
action = self.predict(state)
|
||||
else:
|
||||
action = random.randrange(self.action_dim)
|
||||
return action
|
||||
```
|
||||
|
||||
这里使用e-greedy策略,即设置一个参数epsilon,如果生成的随机数大于epsilon,就根据网络预测的选择action,否则还是随机选择action,这个epsilon是会逐渐减小的,可以使用线性或者指数减小的方式,但不会减小到零,这样在训练稳定时还能保持一定的探索,这部分可以学习探索与利用(exploration and exploition)相关知识。
|
||||
|
||||
上面讲到的预测函数其实就是根据state选取q值最大的action,如下:
|
||||
|
||||
```python
|
||||
def predict(self,state):
|
||||
with torch.no_grad():
|
||||
state = torch.tensor([state], device=self.device, dtype=torch.float32)
|
||||
q_values = self.policy_net(state)
|
||||
action = q_values.max(1)[1].item()
|
||||
```
|
||||
|
||||
然后是更新函数了:
|
||||
|
||||
```python
|
||||
def update(self):
|
||||
|
||||
if len(self.memory) < self.batch_size:
|
||||
return
|
||||
# 从memory中随机采样transition
|
||||
state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(
|
||||
self.batch_size)
|
||||
'''转为张量
|
||||
例如tensor([[-4.5543e-02, -2.3910e-01, 1.8344e-02, 2.3158e-01],...,[-1.8615e-02, -2.3921e-01, -1.1791e-02, 2.3400e-01]])'''
|
||||
state_batch = torch.tensor(
|
||||
state_batch, device=self.device, dtype=torch.float)
|
||||
action_batch = torch.tensor(action_batch, device=self.device).unsqueeze(
|
||||
1) # 例如tensor([[1],...,[0]])
|
||||
reward_batch = torch.tensor(
|
||||
reward_batch, device=self.device, dtype=torch.float) # tensor([1., 1.,...,1])
|
||||
next_state_batch = torch.tensor(
|
||||
next_state_batch, device=self.device, dtype=torch.float)
|
||||
done_batch = torch.tensor(np.float32(
|
||||
done_batch), device=self.device)
|
||||
|
||||
'''计算当前(s_t,a)对应的Q(s_t, a)'''
|
||||
'''torch.gather:对于a=torch.Tensor([[1,2],[3,4]]),那么a.gather(1,torch.Tensor([[0],[1]]))=torch.Tensor([[1],[3]])'''
|
||||
q_values = self.policy_net(state_batch).gather(
|
||||
dim=1, index=action_batch) # 等价于self.forward
|
||||
# 计算所有next states的V(s_{t+1}),即通过target_net中选取reward最大的对应states
|
||||
next_q_values = self.target_net(next_state_batch).max(
|
||||
1)[0].detach() # 比如tensor([ 0.0060, -0.0171,...,])
|
||||
# 计算 expected_q_value
|
||||
# 对于终止状态,此时done_batch[0]=1, 对应的expected_q_value等于reward
|
||||
expected_q_values = reward_batch + \
|
||||
self.gamma * next_q_values * (1-done_batch)
|
||||
# self.loss = F.smooth_l1_loss(q_values,expected_q_values.unsqueeze(1)) # 计算 Huber loss
|
||||
loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1)) # 计算 均方误差loss
|
||||
# 优化模型
|
||||
self.optimizer.zero_grad() # zero_grad清除上一步所有旧的gradients from the last step
|
||||
# loss.backward()使用backpropagation计算loss相对于所有parameters(需要gradients)的微分
|
||||
loss.backward()
|
||||
# for param in self.policy_net.parameters(): # clip防止梯度爆炸
|
||||
# param.grad.data.clamp_(-1, 1)
|
||||
self.optimizer.step() # 更新模型
|
||||
```
|
||||
|
||||
更新遵循伪代码的以下部分:
|
||||
|
||||
<img src="assets/image-20210507162813393.png" alt="image-20210507162813393" style="zoom:50%;" />
|
||||
|
||||
首先从replay buffer中选取一个batch的数据,计算loss,然后进行minibatch SGD。
|
||||
|
||||
然后是保存与加载模型的部分,如下:
|
||||
|
||||
```python
|
||||
def save(self, path):
|
||||
torch.save(self.target_net.state_dict(), path+'dqn_checkpoint.pth')
|
||||
def load(self, path):
|
||||
self.target_net.load_state_dict(torch.load(path+'dqn_checkpoint.pth'))
|
||||
for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()):
|
||||
param.data.copy_(target_param.data)
|
||||
```
|
||||
|
||||
|
||||
|
||||
### 实验结果
|
||||
|
||||
训练结果如下:
|
||||
|
||||
<img src="assets/train_rewards_curve.png" alt="train_rewards_curve" style="zoom: 67%;" />
|
||||
|
||||
<img src="assets/eval_rewards_curve.png" alt="eval_rewards_curve" style="zoom:67%;" />
|
||||
|
||||
## 参考
|
||||
|
||||
|
||||
@@ -5,7 +5,7 @@
|
||||
@Email: johnjim0816@gmail.com
|
||||
@Date: 2020-06-12 00:50:49
|
||||
@LastEditor: John
|
||||
LastEditTime: 2021-04-29 22:19:18
|
||||
LastEditTime: 2021-05-07 16:30:05
|
||||
@Discription:
|
||||
@Environment: python 3.7.7
|
||||
'''
|
||||
@@ -35,15 +35,13 @@ class DQN:
|
||||
(cfg.epsilon_start - cfg.epsilon_end) * \
|
||||
math.exp(-1. * frame_idx / cfg.epsilon_decay)
|
||||
self.batch_size = cfg.batch_size
|
||||
self.policy_net = MLP(state_dim, action_dim,
|
||||
hidden_dim=cfg.hidden_dim).to(self.device)
|
||||
self.target_net = MLP(state_dim, action_dim,
|
||||
hidden_dim=cfg.hidden_dim).to(self.device)
|
||||
for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()):
|
||||
self.policy_net = MLP(state_dim, action_dim,hidden_dim=cfg.hidden_dim).to(self.device)
|
||||
self.target_net = MLP(state_dim, action_dim,hidden_dim=cfg.hidden_dim).to(self.device)
|
||||
for target_param, param in zip(self.target_net.parameters(),self.policy_net.parameters()): # copy params from policy net
|
||||
target_param.data.copy_(param.data)
|
||||
self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg.lr)
|
||||
self.loss = 0
|
||||
self.memory = ReplayBuffer(cfg.memory_capacity)
|
||||
|
||||
|
||||
def choose_action(self, state):
|
||||
'''选择动作
|
||||
@@ -92,11 +90,11 @@ class DQN:
|
||||
expected_q_values = reward_batch + \
|
||||
self.gamma * next_q_values * (1-done_batch)
|
||||
# self.loss = F.smooth_l1_loss(q_values,expected_q_values.unsqueeze(1)) # 计算 Huber loss
|
||||
self.loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1)) # 计算 均方误差loss
|
||||
loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1)) # 计算 均方误差loss
|
||||
# 优化模型
|
||||
self.optimizer.zero_grad() # zero_grad清除上一步所有旧的gradients from the last step
|
||||
# loss.backward()使用backpropagation计算loss相对于所有parameters(需要gradients)的微分
|
||||
self.loss.backward()
|
||||
loss.backward()
|
||||
# for param in self.policy_net.parameters(): # clip防止梯度爆炸
|
||||
# param.grad.data.clamp_(-1, 1)
|
||||
self.optimizer.step() # 更新模型
|
||||
|
||||
BIN
codes/DQN/assets/eval_rewards_curve.png
Normal file
|
After Width: | Height: | Size: 36 KiB |
BIN
codes/DQN/assets/image-20210507162813393.png
Normal file
|
After Width: | Height: | Size: 76 KiB |
BIN
codes/DQN/assets/train_rewards_curve.png
Normal file
|
After Width: | Height: | Size: 37 KiB |
|
Before Width: | Height: | Size: 56 KiB |
|
Before Width: | Height: | Size: 67 KiB |
|
After Width: | Height: | Size: 24 KiB |
|
After Width: | Height: | Size: 25 KiB |
270
codes/DQN/task0_train.ipynb
Normal file
@@ -5,7 +5,7 @@
|
||||
@Email: johnjim0816@gmail.com
|
||||
@Date: 2020-06-12 00:48:57
|
||||
@LastEditor: John
|
||||
LastEditTime: 2021-04-29 22:23:38
|
||||
LastEditTime: 2021-05-05 16:49:15
|
||||
@Discription:
|
||||
@Environment: python 3.7.7
|
||||
'''
|
||||
@@ -14,20 +14,17 @@ curr_path = os.path.dirname(__file__)
|
||||
parent_path = os.path.dirname(curr_path)
|
||||
sys.path.append(parent_path) # add current terminal path to sys.path
|
||||
|
||||
import datetime
|
||||
import torch
|
||||
import gym
|
||||
import torch
|
||||
import datetime
|
||||
|
||||
from common.utils import save_results, make_dir, del_empty_dir
|
||||
from common.utils import save_results, make_dir
|
||||
from common.plot import plot_rewards
|
||||
from DQN.agent import DQN
|
||||
|
||||
|
||||
|
||||
curr_time = datetime.datetime.now().strftime(
|
||||
"%Y%m%d-%H%M%S") # obtain current time
|
||||
|
||||
|
||||
class DQNConfig:
|
||||
def __init__(self):
|
||||
self.algo = "DQN" # name of algo
|
||||
@@ -35,21 +32,21 @@ class DQNConfig:
|
||||
self.result_path = curr_path+"/outputs/" + self.env + \
|
||||
'/'+curr_time+'/results/' # path to save results
|
||||
self.model_path = curr_path+"/outputs/" + self.env + \
|
||||
'/'+curr_time+'/models/' # path to save results
|
||||
self.train_eps = 300 # 训练的episode数目
|
||||
'/'+curr_time+'/models/' # path to save models
|
||||
self.train_eps = 300 # max trainng episodes
|
||||
self.eval_eps = 50 # number of episodes for evaluating
|
||||
self.gamma = 0.95
|
||||
self.epsilon_start = 0.90 # e-greedy策略的初始epsilon
|
||||
self.epsilon_start = 0.90 # start epsilon of e-greedy policy
|
||||
self.epsilon_end = 0.01
|
||||
self.epsilon_decay = 500
|
||||
self.lr = 0.0001 # learning rate
|
||||
self.memory_capacity = 100000 # Replay Memory容量
|
||||
self.memory_capacity = 100000 # capacity of Replay Memory
|
||||
self.batch_size = 64
|
||||
self.target_update = 2 # target net的更新频率
|
||||
self.target_update = 4 # update frequency of target net
|
||||
self.device = torch.device(
|
||||
"cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
|
||||
self.hidden_dim = 256 # 神经网络隐藏层维度
|
||||
|
||||
"cuda" if torch.cuda.is_available() else "cpu") # check gpu
|
||||
self.hidden_dim = 256 # hidden size of net
|
||||
|
||||
def env_agent_config(cfg,seed=1):
|
||||
env = gym.make(cfg.env)
|
||||
env.seed(seed)
|
||||
@@ -63,7 +60,7 @@ def train(cfg, env, agent):
|
||||
print(f'Env:{cfg.env}, Algorithm:{cfg.algo}, Device:{cfg.device}')
|
||||
rewards = []
|
||||
ma_rewards = [] # moveing average reward
|
||||
for i_episode in range(cfg.train_eps):
|
||||
for i_ep in range(cfg.train_eps):
|
||||
state = env.reset()
|
||||
done = False
|
||||
ep_reward = 0
|
||||
@@ -76,11 +73,12 @@ def train(cfg, env, agent):
|
||||
agent.update()
|
||||
if done:
|
||||
break
|
||||
if i_episode % cfg.target_update == 0:
|
||||
if (i_ep+1) % cfg.target_update == 0:
|
||||
agent.target_net.load_state_dict(agent.policy_net.state_dict())
|
||||
print('Episode:{}/{}, Reward:{}'.format(i_episode+1, cfg.train_eps, ep_reward))
|
||||
if (i_ep+1)%10 == 0:
|
||||
print('Episode:{}/{}, Reward:{}'.format(i_ep+1, cfg.train_eps, ep_reward))
|
||||
rewards.append(ep_reward)
|
||||
# 计算滑动窗口的reward
|
||||
# save ma rewards
|
||||
if ma_rewards:
|
||||
ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
|
||||
else:
|
||||
@@ -89,15 +87,17 @@ def train(cfg, env, agent):
|
||||
return rewards, ma_rewards
|
||||
|
||||
def eval(cfg,env,agent):
|
||||
rewards = [] # 记录所有episode的reward
|
||||
ma_rewards = [] # 滑动平均的reward
|
||||
print('Start to eval !')
|
||||
print(f'Env:{cfg.env}, Algorithm:{cfg.algo}, Device:{cfg.device}')
|
||||
rewards = []
|
||||
ma_rewards = [] # moving average rewards
|
||||
for i_ep in range(cfg.eval_eps):
|
||||
ep_reward = 0 # 记录每个episode的reward
|
||||
state = env.reset() # 重置环境, 重新开一局(即开始新的一个episode)
|
||||
ep_reward = 0 # reward per episode
|
||||
state = env.reset()
|
||||
while True:
|
||||
action = agent.predict(state) # 根据算法选择一个动作
|
||||
next_state, reward, done, _ = env.step(action) # 与环境进行一个交互
|
||||
state = next_state # 存储上一个观察值
|
||||
action = agent.predict(state)
|
||||
next_state, reward, done, _ = env.step(action)
|
||||
state = next_state
|
||||
ep_reward += reward
|
||||
if done:
|
||||
break
|
||||
@@ -106,11 +106,15 @@ def eval(cfg,env,agent):
|
||||
ma_rewards.append(ma_rewards[-1]*0.9+ep_reward*0.1)
|
||||
else:
|
||||
ma_rewards.append(ep_reward)
|
||||
print(f"Episode:{i_ep+1}/{cfg.eval_eps}, reward:{ep_reward:.1f}")
|
||||
if (i_ep+1)%10 == 10:
|
||||
print(f"Episode:{i_ep+1}/{cfg.eval_eps}, reward:{ep_reward:.1f}")
|
||||
print('Complete evaling!')
|
||||
return rewards,ma_rewards
|
||||
|
||||
if __name__ == "__main__":
|
||||
cfg = DQNConfig()
|
||||
|
||||
# train
|
||||
env,agent = env_agent_config(cfg,seed=1)
|
||||
rewards, ma_rewards = train(cfg, env, agent)
|
||||
make_dir(cfg.result_path, cfg.model_path)
|
||||
@@ -118,7 +122,7 @@ if __name__ == "__main__":
|
||||
save_results(rewards, ma_rewards, tag='train', path=cfg.result_path)
|
||||
plot_rewards(rewards, ma_rewards, tag="train",
|
||||
algo=cfg.algo, path=cfg.result_path)
|
||||
|
||||
# eval
|
||||
env,agent = env_agent_config(cfg,seed=10)
|
||||
agent.load(path=cfg.model_path)
|
||||
rewards,ma_rewards = eval(cfg,env,agent)
|
||||
|
||||