update codes

codes/DQN-series/DQN/README.md

@@ -0,0 +1,218 @@
+# DQN
+
+## 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。
+
+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。
+
+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)。
+
+https://blog.csdn.net/JohnJim0/article/details/109557173)
+
+## 伪代码
+
+<img src="assets/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L0pvaG5KaW0w,size_16,color_FFFFFF,t_70.png" alt="img" style="zoom:50%;" />
+
+## 代码实现
+
+### 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
+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        
+```
+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
+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
+
+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
+
+class FCN(nn.Module):
+    def __init__(self, state_dim=4, action_dim=18):
+        """ 初始化q网络，为全连接网络
+            state_dim: 输入的feature即环境的state数目
+            action_dim: 输出的action总个数
+        """
+        super(FCN, self).__init__()
+        self.fc1 = nn.Linear(state_dim, 128) # 输入层
+        self.fc2 = nn.Linear(128, 128) # 隐藏层
+        self.fc3 = nn.Linear(128, action_dim) # 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_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```)：
+
+```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)
+```
+
+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
+
+[with torch.no_grad()](https://www.jianshu.com/p/1cea017f5d11)
+

codes/DQN-series/DQN/agent.py

@@ -0,0 +1,84 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+@Author: John
+@Email: johnjim0816@gmail.com
+@Date: 2020-06-12 00:50:49
+@LastEditor: John
+LastEditTime: 2021-09-15 13:35:36
+@Discription: 
+@Environment: python 3.7.7
+'''
+'''off-policy
+'''
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import random
+import math
+import numpy as np
+from common.memory import ReplayBuffer
+from common.model import MLP
+class DQN:
+    def __init__(self, n_states, n_actions, cfg):
+
+        self.n_actions = n_actions  # 总的动作个数
+        self.device = cfg.device  # 设备，cpu或gpu等
+        self.gamma = cfg.gamma  # 奖励的折扣因子
+        # e-greedy策略相关参数
+        self.frame_idx = 0  # 用于epsilon的衰减计数
+        self.epsilon = lambda frame_idx: cfg.epsilon_end + \
+            (cfg.epsilon_start - cfg.epsilon_end) * \
+            math.exp(-1. * frame_idx / cfg.epsilon_decay)
+        self.batch_size = cfg.batch_size
+        self.policy_net = MLP(n_states, n_actions,hidden_dim=cfg.hidden_dim).to(self.device)
+        self.target_net = MLP(n_states, n_actions,hidden_dim=cfg.hidden_dim).to(self.device)
+        for target_param, param in zip(self.target_net.parameters(),self.policy_net.parameters()): # 复制参数到目标网路targe_net
+            target_param.data.copy_(param.data)
+        self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg.lr) # 优化器
+        self.memory = ReplayBuffer(cfg.memory_capacity) # 经验回放
+
+    def choose_action(self, state):
+        ''' 选择动作
+        '''
+        self.frame_idx += 1
+        if random.random() > self.epsilon(self.frame_idx):
+            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() # 选择Q值最大的动作
+        else:
+            action = random.randrange(self.n_actions)
+        return action
+    def update(self):
+        if len(self.memory) < self.batch_size: # 当memory中不满足一个批量时，不更新策略
+            return
+        # 从经验回放中(replay memory)中随机采样一个批量的转移(transition)
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(
+            self.batch_size)
+        # 转为张量
+        state_batch = torch.tensor(state_batch, device=self.device, dtype=torch.float)
+        action_batch = torch.tensor(action_batch, device=self.device).unsqueeze(1)  
+        reward_batch = torch.tensor(reward_batch, device=self.device, dtype=torch.float)  
+        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)
+        q_values = self.policy_net(state_batch).gather(dim=1, index=action_batch) # 计算当前状态(s_t,a)对应的Q(s_t, a)
+        next_q_values = self.target_net(next_state_batch).max(1)[0].detach() # 计算下一时刻的状态(s_t_,a)对应的Q值
+        # 计算期望的Q值，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)
+        loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1))  # 计算均方根损失
+        # 优化更新模型
+        self.optimizer.zero_grad()  
+        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() 
+
+    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)

codes/DQN-series/DQN/task0_train.py

@@ -0,0 +1,136 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+@Author: John
+@Email: johnjim0816@gmail.com
+@Date: 2020-06-12 00:48:57
+@LastEditor: John
+LastEditTime: 2021-09-15 15:34:13
+@Discription: 
+@Environment: python 3.7.7
+'''
+import sys,os
+curr_path = os.path.dirname(os.path.abspath(__file__)) # 当前文件所在绝对路径
+parent_path = os.path.dirname(curr_path) # 父路径
+sys.path.append(parent_path) # 添加父路径到系统路径sys.path
+
+import gym
+import torch
+import datetime
+
+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")  # 获取当前时间
+class DQNConfig:
+    def __init__(self):
+        self.algo = "DQN"  # 算法名称
+        self.env = 'CartPole-v0' # 环境名称
+        self.train_eps = 200 # 训练的回合数
+        self.eval_eps = 30 # 测试的回合数
+        self.gamma = 0.95 # 强化学习中的折扣因子
+        self.epsilon_start = 0.90 # e-greedy策略中初始epsilon
+        self.epsilon_end = 0.01 # e-greedy策略中的终止epsilon
+        self.epsilon_decay = 500 # e-greedy策略中epsilon的衰减率
+        self.lr = 0.0001  # 学习率
+        self.memory_capacity = 100000  # 经验回放的容量
+        self.batch_size = 64 # mini-batch SGD中的批量大小
+        self.target_update = 4 # 目标网络的更新频率
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")  # 检测GPU
+        self.hidden_dim = 256  # 网络隐藏层
+class PlotConfig:
+    def __init__(self) -> None:
+        self.algo = "DQN"  # 算法名称
+        self.env = 'CartPole-v0' # 环境名称
+        self.result_path = curr_path+"/outputs/" + self.env + \
+            '/'+curr_time+'/results/'  # 保存结果的路径
+        self.model_path = curr_path+"/outputs/" + self.env + \
+            '/'+curr_time+'/models/'  # 保存模型的路径
+        self.save = True # 是否保存图片
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")  # 检测GPU
+
+def env_agent_config(cfg,seed=1):
+    ''' 创建环境和智能体
+    '''
+    env = gym.make(cfg.env)  # 创建环境
+    env.seed(seed) # 设置随机种子
+    n_states = env.observation_space.shape[0] # 状态数
+    n_actions = env.action_space.n # 动作数
+    agent = DQN(n_states,n_actions,cfg) # 创建智能体
+    return env,agent
+    
+def train(cfg, env, agent):
+    ''' 训练
+    '''
+    print('开始训练!')
+    print(f'环境：{cfg.env}, 算法：{cfg.algo}, 设备：{cfg.device}')
+    rewards = [] # 记录所有回合的奖励
+    ma_rewards = []  # 记录所有回合的滑动平均奖励
+    for i_ep in range(cfg.train_eps):
+        ep_reward = 0 # 记录一回合内的奖励
+        state = env.reset() # 重置环境，返回初始状态
+        while True:
+            action = agent.choose_action(state) # 选择动作
+            next_state, reward, done, _ = env.step(action) # 更新环境，返回transition
+            agent.memory.push(state, action, reward, next_state, done) # 保存transition
+            state = next_state # 更新下一个状态
+            agent.update() # 更新智能体
+            ep_reward += reward # 累加奖励
+            if done:
+                break
+        if (i_ep+1) % cfg.target_update == 0: # 智能体目标网络更新
+            agent.target_net.load_state_dict(agent.policy_net.state_dict())
+        if (i_ep+1)%10 == 0: 
+            print('回合：{}/{}, 奖励：{}'.format(i_ep+1, cfg.train_eps, ep_reward))
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
+        else:
+            ma_rewards.append(ep_reward)
+    print('完成训练！')
+    return rewards, ma_rewards
+
+def eval(cfg,env,agent):
+    print('开始测试!')
+    print(f'环境：{cfg.env}, 算法：{cfg.algo}, 设备：{cfg.device}')
+    # 由于测试不需要使用epsilon-greedy策略，所以相应的值设置为0
+    cfg.epsilon_start = 0.0 # e-greedy策略中初始epsilon
+    cfg.epsilon_end = 0.0 # e-greedy策略中的终止epsilon
+    rewards = [] # 记录所有回合的奖励
+    ma_rewards = []  # 记录所有回合的滑动平均奖励
+    for i_ep in range(cfg.eval_eps):
+        ep_reward = 0 # 记录一回合内的奖励
+        state = env.reset() # 重置环境，返回初始状态
+        while True:
+            action = agent.choose_action(state) # 选择动作
+            next_state, reward, done, _ = env.step(action) # 更新环境，返回transition
+            state = next_state # 更新下一个状态
+            ep_reward += reward # 累加奖励
+            if done:
+                break
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(ma_rewards[-1]*0.9+ep_reward*0.1)
+        else:
+            ma_rewards.append(ep_reward)
+        print(f"回合：{i_ep+1}/{cfg.eval_eps}, 奖励：{ep_reward:.1f}")
+    print('完成测试！')
+    return rewards,ma_rewards
+
+if __name__ == "__main__":
+    cfg = DQNConfig()
+    plot_cfg = PlotConfig()
+    # 训练
+    env,agent = env_agent_config(cfg,seed=1)
+    rewards, ma_rewards = train(cfg, env, agent)
+    make_dir(plot_cfg.result_path, plot_cfg.model_path) # 创建保存结果和模型路径的文件夹
+    agent.save(path=plot_cfg.model_path) # 保存模型
+    save_results(rewards, ma_rewards, tag='train', path=plot_cfg.result_path) # 保存结果
+    plot_rewards(rewards, ma_rewards, plot_cfg, tag="train") # 画出结果
+    # 测试
+    env,agent = env_agent_config(cfg,seed=10)
+    agent.load(path=plot_cfg.model_path) # 导入模型
+    rewards,ma_rewards = eval(cfg,env,agent)
+    save_results(rewards,ma_rewards,tag='eval',path=plot_cfg.result_path) # 保存结果
+    plot_rewards(rewards,ma_rewards, plot_cfg, tag="eval") # 画出结果

codes/DQN-series/DoubleDQN/README.md

@@ -0,0 +1,39 @@
+食用本篇之前，需要有DQN算法的基础，参考[DQN算法实战](../DQN)。
+
+## 原理简介
+
+Double-DQN是2016年提出的算法，灵感源自2010年的Double-Qlearning，可参考论文[Deep Reinforcement Learning with Double Q-learning](https://arxiv.org/abs/1509.06461)。
+跟Nature DQN一样，Double-DQN也用了两个网络，一个当前网络(对应用$Q$表示)，一个目标网络(对应一般用$Q'$表示，为方便区分，以下用$Q_{tar}$代替)。我们先回忆一下，对于非终止状态，目标$Q_{tar}$值计算如下
+![在这里插入图片描述](assets/20201222145725907.png)
+
+而在Double-DQN中，不再是直接从目标$Q_{tar}$网络中选择各个动作中的最大$Q_{tar}$值，而是先从当前$Q$网络选择$Q$值最大对应的动作，然后代入到目标网络中计算对应的值：
+![在这里插入图片描述](assets/20201222150225327.png)
+Double-DQN的好处是Nature DQN中使用max虽然可以快速让Q值向可能的优化目标靠拢，但是很容易过犹不及，导致过度估计(Over Estimation)，所谓过度估计就是最终我们得到的算法模型有很大的偏差(bias)。为了解决这个问题， DDQN通过解耦目标Q值动作的选择和目标Q值的计算这两步，来达到消除过度估计的问题，感兴趣可以阅读原论文。
+
+伪代码如下：
+![在这里插入图片描述](assets/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L0pvaG5KaW0w,size_16,color_FFFFFF,t_70.png)
+当然也可以两个网络可以同时为当前网络和目标网络，如下：
+![在这里插入图片描述](assets/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L0pvaG5KaW0w,size_16,color_FFFFFF,t_70-20210328110837146.png)
+或者这样更好理解如何同时为当前网络和目标网络：
+![在这里插入图片描述](assets/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L0pvaG5KaW0w,size_16,color_FFFFFF,t_70-20210328110837157.png)
+
+## 代码实战
+完整程序见[github](https://github.com/JohnJim0816/reinforcement-learning-tutorials/tree/master/DoubleDQN)。结合上面的原理，其实Double DQN改进来很简单，基本只需要在```update```中修改几行代码，如下：
+```python
+'''以下是Nature DQN的q_target计算方式
+next_q_state_value = self.target_net(
+next_state_batch).max(1)[0].detach()  # # 计算所有next states的Q'(s_{t+1})的最大值，Q'为目标网络的q函数,比如tensor([ 0.0060, -0.0171,...,])
+#计算 q_target
+#对于终止状态，此时done_batch[0]=1, 对应的expected_q_value等于reward
+q_target = reward_batch + self.gamma * next_q_state_value * (1-done_batch[0])
+'''
+'''以下是Double DQNq_target计算方式，与NatureDQN稍有不同'''
+next_target_values = self.target_net(
+next_state_batch)
+#选出Q(s_t‘, a)对应的action，代入到next_target_values获得target net对应的next_q_value，即Q’(s_t|a=argmax Q(s_t‘, a))
+next_target_q_value = next_target_values.gather(1, torch.max(next_q_values, 1)[1].unsqueeze(1)).squeeze(1)
+q_target = reward_batch + self.gamma * next_target_q_value * (1-done_batch[0])
+```
+reward变化结果如下：
+![在这里插入图片描述](assets/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L0pvaG5KaW0w,size_16,color_FFFFFF,t_70-20210328110837128.png)
+其中下边蓝色和红色分别表示Double DQN和Nature DQN在训练中的reward变化图，而上面蓝色和绿色则表示Double DQN和Nature DQN在测试中的reward变化图。

codes/DQN-series/DoubleDQN/agent.py

@@ -0,0 +1,123 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+@Author: John
+@Email: johnjim0816@gmail.com
+@Date: 2020-06-12 00:50:49
+@LastEditor: John
+LastEditTime: 2021-05-04 22:28:06
+@Discription: 
+@Environment: python 3.7.7
+'''
+'''off-policy
+'''
+
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import random
+import math
+import numpy as np
+from common.memory import ReplayBuffer
+from common.model import MLP
+class DoubleDQN:
+    def __init__(self, state_dim, action_dim, cfg):
+        
+        self.action_dim = action_dim  # 总的动作个数
+        self.device = cfg.device  # 设备，cpu或gpu等
+        self.gamma = cfg.gamma
+        # e-greedy策略相关参数
+        self.actions_count = 0
+        self.epsilon_start = cfg.epsilon_start
+        self.epsilon_end = cfg.epsilon_end
+        self.epsilon_decay = 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)
+        # target_net copy from policy_net
+        for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()):
+            target_param.data.copy_(param.data)
+        # self.target_net.eval()  # 不启用 BatchNormalization 和 Dropout
+        # 可查parameters()与state_dict()的区别，前者require_grad=True
+        self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg.lr)
+        self.loss = 0
+        self.memory = ReplayBuffer(cfg.memory_capacity)
+    def predict(self,state):
+        with torch.no_grad():
+                # 先转为张量便于丢给神经网络,state元素数据原本为float64
+                # 注意state=torch.tensor(state).unsqueeze(0)跟state=torch.tensor([state])等价
+                state = torch.tensor(
+                    [state], device=self.device, dtype=torch.float32)
+                # 如tensor([[-0.0798, -0.0079]], grad_fn=<AddmmBackward>)
+                q_value = self.policy_net(state)
+                # tensor.max(1)返回每行的最大值以及对应的下标，
+                # 如torch.return_types.max(values=tensor([10.3587]),indices=tensor([0]))
+                # 所以tensor.max(1)[1]返回最大值对应的下标，即action
+                action = q_value.max(1)[1].item()  
+        return action
+    def choose_action(self, state):
+        '''选择动作
+        '''
+        self.actions_count += 1
+        self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \
+            math.exp(-1. * self.actions_count / self.epsilon_decay)
+        if random.random() > self.epsilon:
+            action  = self.predict(state)
+        else:
+            action = random.randrange(self.action_dim)
+        return action
+    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)
+        # convert to tensor
+        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)  # 将bool转为float然后转为张量
+        # 计算当前(s_t,a)对应的Q(s_t, a)
+        q_values = self.policy_net(state_batch) 
+        next_q_values = self.policy_net(next_state_batch)
+        # 代入当前选择的action，得到Q(s_t|a=a_t)
+        q_value = q_values.gather(dim=1, index=action_batch)
+        '''以下是Nature DQN的q_target计算方式
+        # 计算所有next states的Q'(s_{t+1})的最大值，Q'为目标网络的q函数
+        next_q_state_value = self.target_net(
+            next_state_batch).max(1)[0].detach()  # 比如tensor([ 0.0060, -0.0171,...,])
+        # 计算 q_target
+        # 对于终止状态，此时done_batch[0]=1, 对应的expected_q_value等于reward
+        q_target = reward_batch + self.gamma * next_q_state_value * (1-done_batch[0])
+        '''
+        '''以下是Double DQN q_target计算方式，与NatureDQN稍有不同'''
+        next_target_values = self.target_net(
+            next_state_batch)
+        # 选出Q(s_t‘, a)对应的action，代入到next_target_values获得target net对应的next_q_value，即Q’(s_t|a=argmax Q(s_t‘, a))
+        next_target_q_value = next_target_values.gather(1, torch.max(next_q_values, 1)[1].unsqueeze(1)).squeeze(1)
+        q_target = reward_batch + self.gamma * next_target_q_value * (1-done_batch)
+        self.loss = nn.MSELoss()(q_value, q_target.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()
+        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
+
+    def save(self,path):
+        torch.save(self.target_net.state_dict(), path+'checkpoint.pth')
+
+    def load(self,path):
+        self.target_net.load_state_dict(torch.load(path+'checkpoint.pth'))  
+        for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()):
+            param.data.copy_(target_param.data)  

codes/DQN-series/DoubleDQN/task0_train.ipynb

@@ -0,0 +1,194 @@
+{
+ "metadata": {
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.7.10"
+  },
+  "orig_nbformat": 2,
+  "kernelspec": {
+   "name": "python3710jvsc74a57bd0366e1054dee9d4501b0eb8f87335afd3c67fc62db6ee611bbc7f8f5a1fefe232",
+   "display_name": "Python 3.7.10 64-bit ('py37': conda)"
+  },
+  "metadata": {
+   "interpreter": {
+    "hash": "366e1054dee9d4501b0eb8f87335afd3c67fc62db6ee611bbc7f8f5a1fefe232"
+   }
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2,
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import sys\n",
+    "from pathlib import Path\n",
+    "curr_path = str(Path().absolute())\n",
+    "parent_path = str(Path().absolute().parent)\n",
+    "sys.path.append(parent_path) # add current terminal path to sys.path"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import gym\n",
+    "import torch\n",
+    "import datetime\n",
+    "from DoubleDQN.agent import DoubleDQN\n",
+    "from common.plot import plot_rewards\n",
+    "from common.utils import save_results, make_dir\n",
+    "\n",
+    "curr_time = datetime.datetime.now().strftime(\n",
+    "    \"%Y%m%d-%H%M%S\")  # obtain current time"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "class DoubleDQNConfig:\n",
+    "    def __init__(self):\n",
+    "        self.algo = \"DoubleDQN\" # name of algo\n",
+    "        self.env = 'CartPole-v0'  # env name\n",
+    "        self.result_path = curr_path+\"/outputs/\" + self.env + \\\n",
+    "            '/'+curr_time+'/results/'  # path to save results\n",
+    "        self.model_path = curr_path+\"/outputs/\" + self.env + \\\n",
+    "            '/'+curr_time+'/models/'  # path to save models\n",
+    "        self.train_eps = 200 # max tranng episodes\n",
+    "        self.eval_eps = 50 # max evaling episodes\n",
+    "        self.gamma = 0.95\n",
+    "        self.epsilon_start = 1 # start epsilon of e-greedy policy\n",
+    "        self.epsilon_end = 0.01 \n",
+    "        self.epsilon_decay = 500\n",
+    "        self.lr = 0.001 # learning rate\n",
+    "        self.memory_capacity = 100000 # capacity of Replay Memory\n",
+    "        self.batch_size = 64\n",
+    "        self.target_update = 2 # update frequency of target net\n",
+    "        self.device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\") # check gpu\n",
+    "        self.hidden_dim = 256 # hidden size of net"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def env_agent_config(cfg,seed=1):\n",
+    "    env = gym.make(cfg.env)  \n",
+    "    env.seed(seed)\n",
+    "    state_dim = env.observation_space.shape[0]\n",
+    "    action_dim = env.action_space.n\n",
+    "    agent = DoubleDQN(state_dim,action_dim,cfg)\n",
+    "    return env,agent"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def train(cfg,env,agent):\n",
+    "    print('Start to train !')\n",
+    "    rewards,ma_rewards = [],[]\n",
+    "    for i_ep in range(cfg.train_eps):\n",
+    "        state = env.reset() \n",
+    "        ep_reward = 0\n",
+    "        while True:\n",
+    "            action = agent.choose_action(state) \n",
+    "            next_state, reward, done, _ = env.step(action)\n",
+    "            ep_reward += reward\n",
+    "            agent.memory.push(state, action, reward, next_state, done) \n",
+    "            state = next_state \n",
+    "            agent.update() \n",
+    "            if done:\n",
+    "                break\n",
+    "        if i_ep % cfg.target_update == 0:\n",
+    "            agent.target_net.load_state_dict(agent.policy_net.state_dict())\n",
+    "        if (i_ep+1)%10 == 0:\n",
+    "            print(f'Episode:{i_ep+1}/{cfg.train_eps}, Reward:{ep_reward}')\n",
+    "        rewards.append(ep_reward)\n",
+    "        if ma_rewards:\n",
+    "            ma_rewards.append(\n",
+    "                0.9*ma_rewards[-1]+0.1*ep_reward)\n",
+    "        else:\n",
+    "            ma_rewards.append(ep_reward)   \n",
+    "    print('Complete training！')\n",
+    "    return rewards,ma_rewards"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def eval(cfg,env,agent):\n",
+    "    print('Start to eval !')\n",
+    "    print(f'Env:{cfg.env}, Algorithm:{cfg.algo}, Device:{cfg.device}')\n",
+    "    rewards = []  \n",
+    "    ma_rewards = []\n",
+    "    for i_ep in range(cfg.eval_eps):\n",
+    "        state = env.reset() \n",
+    "        ep_reward = 0   \n",
+    "        while True:\n",
+    "            action = agent.predict(state)  \n",
+    "            next_state, reward, done, _ = env.step(action)  \n",
+    "            state = next_state  \n",
+    "            ep_reward += reward\n",
+    "            if done:\n",
+    "                break\n",
+    "        rewards.append(ep_reward)\n",
+    "        if ma_rewards:\n",
+    "            ma_rewards.append(ma_rewards[-1]*0.9+ep_reward*0.1)\n",
+    "        else:\n",
+    "            ma_rewards.append(ep_reward)\n",
+    "        print(f\"Episode:{i_ep+1}/{cfg.eval_eps}, reward:{ep_reward:.1f}\")\n",
+    "    print('Complete evaling！')\n",
+    "    return rewards,ma_rewards "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "if __name__ == \"__main__\":\n",
+    "    cfg = DoubleDQNConfig()\n",
+    "    # train\n",
+    "    env,agent = env_agent_config(cfg,seed=1)\n",
+    "    rewards, ma_rewards = train(cfg, env, agent)\n",
+    "    make_dir(cfg.result_path, cfg.model_path)\n",
+    "    agent.save(path=cfg.model_path)\n",
+    "    save_results(rewards, ma_rewards, tag='train', path=cfg.result_path)\n",
+    "    plot_rewards(rewards, ma_rewards, tag=\"train\",\n",
+    "                 algo=cfg.algo, path=cfg.result_path)\n",
+    "\n",
+    "    # eval\n",
+    "    env,agent = env_agent_config(cfg,seed=10)\n",
+    "    agent.load(path=cfg.model_path)\n",
+    "    rewards,ma_rewards = eval(cfg,env,agent)\n",
+    "    save_results(rewards,ma_rewards,tag='eval',path=cfg.result_path)\n",
+    "    plot_rewards(rewards,ma_rewards,tag=\"eval\",env=cfg.env,algo = cfg.algo,path=cfg.result_path)"
+   ]
+  }
+ ]
+}

codes/DQN-series/DoubleDQN/task0_train.py

@@ -0,0 +1,123 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+@Author: John
+@Email: johnjim0816@gmail.com
+@Date: 2020-06-12 00:48:57
+@LastEditor: John
+LastEditTime: 2021-09-10 15:26:05
+@Discription: 
+@Environment: python 3.7.7
+'''
+import sys,os
+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 gym
+import torch
+import datetime
+from DoubleDQN.agent import DoubleDQN
+from common.plot import plot_rewards
+from common.utils import save_results, make_dir
+
+curr_time = datetime.datetime.now().strftime(
+    "%Y%m%d-%H%M%S")  # obtain current time
+
+class DoubleDQNConfig:
+    def __init__(self):
+        self.algo = "DoubleDQN" # name of algo
+        self.env = 'CartPole-v0'  # env name
+        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 models
+        self.train_eps = 200 # max tranng episodes
+        self.eval_eps = 50 # max evaling episodes
+        self.gamma = 0.95
+        self.epsilon_start = 1 # start epsilon of e-greedy policy
+        self.epsilon_end = 0.01 
+        self.epsilon_decay = 500
+        self.lr = 0.001 # learning rate
+        self.memory_capacity = 100000 # capacity of Replay Memory
+        self.batch_size = 64
+        self.target_update = 2 # update frequency of target net
+        self.device = torch.device("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)
+    state_dim = env.observation_space.shape[0]
+    action_dim = env.action_space.n
+    agent = DoubleDQN(state_dim,action_dim,cfg)
+    return env,agent
+    
+def train(cfg,env,agent):
+    print('Start to train !')
+    rewards,ma_rewards = [],[]
+    for i_ep in range(cfg.train_eps):
+        state = env.reset() 
+        ep_reward = 0
+        while True:
+            action = agent.choose_action(state) 
+            next_state, reward, done, _ = env.step(action)
+            ep_reward += reward
+            agent.memory.push(state, action, reward, next_state, done) 
+            state = next_state 
+            agent.update() 
+            if done:
+                break
+        if i_ep % cfg.target_update == 0:
+            agent.target_net.load_state_dict(agent.policy_net.state_dict())
+        print(f'Episode:{i_ep+1}/{cfg.train_eps}, Reward:{ep_reward},Epsilon:{agent.epsilon:.2f}')
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(
+                0.9*ma_rewards[-1]+0.1*ep_reward)
+        else:
+            ma_rewards.append(ep_reward)   
+    print('Complete training！')
+    return rewards,ma_rewards
+
+def eval(cfg,env,agent):
+    print('Start to eval !')
+    print(f'Env:{cfg.env}, Algorithm:{cfg.algo}, Device:{cfg.device}')
+    rewards = []  
+    ma_rewards = []
+    for i_ep in range(cfg.eval_eps):
+        state = env.reset() 
+        ep_reward = 0   
+        while True:
+            action = agent.predict(state)  
+            next_state, reward, done, _ = env.step(action)  
+            state = next_state  
+            ep_reward += reward
+            if done:
+                break
+        rewards.append(ep_reward)
+        if ma_rewards:
+            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}")
+    print('Complete evaling！')
+    return rewards,ma_rewards    
+    
+if __name__ == "__main__":
+    cfg = DoubleDQNConfig()
+    # 训练
+    env,agent = env_agent_config(cfg,seed=1)
+    rewards, ma_rewards = train(cfg, env, agent)
+    make_dir(cfg.result_path, cfg.model_path)
+    agent.save(path=cfg.model_path)
+    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)
+
+    # 测试
+    env,agent = env_agent_config(cfg,seed=10)
+    agent.load(path=cfg.model_path)
+    rewards,ma_rewards = eval(cfg,env,agent)
+    save_results(rewards,ma_rewards,tag='eval',path=cfg.result_path)
+    plot_rewards(rewards,ma_rewards,tag="eval",env=cfg.env,algo = cfg.algo,path=cfg.result_path)

codes/DQN-series/HierarchicalDQN/README.md

@@ -0,0 +1,13 @@
+# Hierarchical DQN
+
+## 原理简介
+
+Hierarchical DQN是一种分层强化学习方法，与DQN相比增加了一个meta controller，
+
+![image-20210331153115575](assets/image-20210331153115575.png)
+
+即学习时，meta controller每次会生成一个goal，然后controller或者说下面的actor就会达到这个goal，直到done为止。这就相当于给agent增加了一个队长，队长擅长制定局部目标，指导agent前行，这样应对一些每回合步数较长或者稀疏奖励的问题会有所帮助。
+
+## 伪代码
+
+![image-20210331153542314](assets/image-20210331153542314.png)

codes/DQN-series/HierarchicalDQN/agent.py

@@ -0,0 +1,115 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2021-03-24 22:18:18
+LastEditor: John
+LastEditTime: 2021-05-04 22:39:34
+Discription: 
+Environment: 
+'''
+import torch
+import torch.nn as nn
+import numpy as np
+import random,math
+import torch.optim as optim
+from common.model import MLP
+from common.memory import ReplayBuffer
+
+class HierarchicalDQN:
+    def __init__(self,state_dim,action_dim,cfg):
+        self.state_dim = state_dim
+        self.action_dim = action_dim
+        self.gamma = cfg.gamma
+        self.device = cfg.device
+        self.batch_size = cfg.batch_size
+        self.frame_idx = 0 
+        self.epsilon = lambda frame_idx: cfg.epsilon_end + (cfg.epsilon_start - cfg.epsilon_end ) * math.exp(-1. * frame_idx / cfg.epsilon_decay)
+        self.policy_net = MLP(2*state_dim, action_dim,cfg.hidden_dim).to(self.device)
+        self.meta_policy_net = MLP(state_dim, state_dim,cfg.hidden_dim).to(self.device)
+        self.optimizer = optim.Adam(self.policy_net.parameters(),lr=cfg.lr)
+        self.meta_optimizer = optim.Adam(self.meta_policy_net.parameters(),lr=cfg.lr)
+        self.memory = ReplayBuffer(cfg.memory_capacity)
+        self.meta_memory = ReplayBuffer(cfg.memory_capacity)
+        self.loss_numpy  = 0
+        self.meta_loss_numpy  = 0
+        self.losses = []
+        self.meta_losses = []
+    def to_onehot(self,x):
+        oh = np.zeros(self.state_dim)
+        oh[x - 1] = 1.
+        return oh
+    def set_goal(self,state):
+        if random.random() > self.epsilon(self.frame_idx):
+            with torch.no_grad():
+                state = torch.tensor(state, device=self.device, dtype=torch.float32).unsqueeze(0)
+                goal = self.meta_policy_net(state).max(1)[1].item() 
+        else:
+            goal = random.randrange(self.state_dim)
+        return goal
+    def choose_action(self,state):
+        self.frame_idx += 1
+        if random.random() > self.epsilon(self.frame_idx):
+            with torch.no_grad():
+                state = torch.tensor(state, device=self.device, dtype=torch.float32).unsqueeze(0)
+                q_value = self.policy_net(state)
+                action = q_value.max(1)[1].item()  
+        else:
+            action = random.randrange(self.action_dim)
+        return action
+    def update(self):
+        self.update_policy()
+        self.update_meta()
+    def update_policy(self): 
+        if self.batch_size > len(self.memory):
+            return
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(self.batch_size)
+        state_batch = torch.tensor(state_batch,device=self.device,dtype=torch.float)
+        action_batch = torch.tensor(action_batch,device=self.device,dtype=torch.int64).unsqueeze(1)  
+        reward_batch = torch.tensor(reward_batch,device=self.device,dtype=torch.float)  
+        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)
+        q_values = self.policy_net(state_batch).gather(dim=1, index=action_batch).squeeze(1)
+        next_state_values = self.policy_net(next_state_batch).max(1)[0].detach()
+        expected_q_values = reward_batch + 0.99 * next_state_values * (1-done_batch)
+        loss = nn.MSELoss()(q_values, expected_q_values) 
+        self.optimizer.zero_grad() 
+        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()  
+        self.loss_numpy = loss.detach().cpu().numpy()
+        self.losses.append(self.loss_numpy)  
+    def update_meta(self):
+        if self.batch_size > len(self.meta_memory):
+            return
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.meta_memory.sample(self.batch_size)
+        state_batch = torch.tensor(state_batch,device=self.device,dtype=torch.float)
+        action_batch = torch.tensor(action_batch,device=self.device,dtype=torch.int64).unsqueeze(1)  
+        reward_batch = torch.tensor(reward_batch,device=self.device,dtype=torch.float)  
+        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)
+        q_values = self.meta_policy_net(state_batch).gather(dim=1, index=action_batch).squeeze(1)
+        next_state_values = self.meta_policy_net(next_state_batch).max(1)[0].detach()
+        expected_q_values = reward_batch + 0.99 * next_state_values * (1-done_batch)
+        meta_loss = nn.MSELoss()(q_values, expected_q_values) 
+        self.meta_optimizer.zero_grad() 
+        meta_loss.backward()
+        for param in self.meta_policy_net.parameters():  # clip防止梯度爆炸
+            param.grad.data.clamp_(-1, 1)
+        self.meta_optimizer.step() 
+        self.meta_loss_numpy = meta_loss.detach().cpu().numpy()
+        self.meta_losses.append(self.meta_loss_numpy)
+
+    def save(self, path):
+        torch.save(self.policy_net.state_dict(), path+'policy_checkpoint.pth')
+        torch.save(self.meta_policy_net.state_dict(), path+'meta_checkpoint.pth')
+
+    def load(self, path):
+        self.policy_net.load_state_dict(torch.load(path+'policy_checkpoint.pth'))
+        self.meta_policy_net.load_state_dict(torch.load(path+'meta_checkpoint.pth'))
+        
+
+        
+        

codes/DQN-series/HierarchicalDQN/task0_train.py

@@ -0,0 +1,146 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2021-03-29 10:37:32
+LastEditor: John
+LastEditTime: 2021-05-04 22:35:56
+Discription: 
+Environment: 
+'''
+
+
+import sys,os
+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 numpy as np
+import torch
+import gym
+
+from common.utils import save_results,make_dir
+from common.plot import plot_rewards
+from HierarchicalDQN.agent import HierarchicalDQN
+
+curr_time = datetime.datetime.now().strftime(
+    "%Y%m%d-%H%M%S")  # obtain current time
+
+class HierarchicalDQNConfig:
+    def __init__(self):
+        self.algo = "H-DQN"  # name of algo
+        self.env = 'CartPole-v0'
+        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 models
+        self.train_eps = 300  # 训练的episode数目
+        self.eval_eps = 50  # 测试的episode数目
+        self.gamma = 0.99
+        self.epsilon_start = 1  # start epsilon of e-greedy policy
+        self.epsilon_end = 0.01
+        self.epsilon_decay = 200
+        self.lr = 0.0001  # learning rate
+        self.memory_capacity = 10000  # Replay Memory capacity
+        self.batch_size = 32
+        self.target_update = 2  # target net的更新频率
+        self.device = torch.device(
+            "cuda" if torch.cuda.is_available() else "cpu")  # 检测gpu
+        self.hidden_dim = 256  # dimension of hidden layer
+
+def env_agent_config(cfg,seed=1):
+    env = gym.make(cfg.env)  
+    env.seed(seed)
+    state_dim = env.observation_space.shape[0]
+    action_dim = env.action_space.n
+    agent = HierarchicalDQN(state_dim,action_dim,cfg)
+    return env,agent
+
+def train(cfg, env, agent):
+    print('Start to train !')
+    print(f'Env:{cfg.env}, Algorithm:{cfg.algo}, Device:{cfg.device}')
+    rewards = []
+    ma_rewards = []  # moveing average reward
+    for i_ep in range(cfg.train_eps):
+        state = env.reset()
+        done = False
+        ep_reward = 0
+        while not done:
+            goal = agent.set_goal(state)
+            onehot_goal = agent.to_onehot(goal)
+            meta_state = state
+            extrinsic_reward = 0
+            while not done and goal != np.argmax(state):
+                goal_state = np.concatenate([state, onehot_goal])
+                action = agent.choose_action(goal_state)
+                next_state, reward, done, _ = env.step(action)
+                ep_reward += reward
+                extrinsic_reward += reward
+                intrinsic_reward = 1.0 if goal == np.argmax(
+                    next_state) else 0.0
+                agent.memory.push(goal_state, action, intrinsic_reward, np.concatenate(
+                    [next_state, onehot_goal]), done)
+                state = next_state
+                agent.update()
+        agent.meta_memory.push(meta_state, goal, extrinsic_reward, state, done)
+        print('Episode:{}/{}, Reward:{}, Loss:{:.2f}, Meta_Loss:{:.2f}'.format(i_ep+1, cfg.train_eps, ep_reward,agent.loss_numpy ,agent.meta_loss_numpy ))
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(
+                0.9*ma_rewards[-1]+0.1*ep_reward)
+        else:
+            ma_rewards.append(ep_reward)
+    print('Complete training！')
+    return rewards, ma_rewards
+
+def eval(cfg, env, agent):
+    print('Start to eval !')
+    print(f'Env:{cfg.env}, Algorithm:{cfg.algo}, Device:{cfg.device}')
+    rewards = []
+    ma_rewards = []  # moveing average reward
+    for i_ep in range(cfg.train_eps):
+        state = env.reset()
+        done = False
+        ep_reward = 0
+        while not done:
+            goal = agent.set_goal(state)
+            onehot_goal = agent.to_onehot(goal)
+            extrinsic_reward = 0
+            while not done and goal != np.argmax(state):
+                goal_state = np.concatenate([state, onehot_goal])
+                action = agent.choose_action(goal_state)
+                next_state, reward, done, _ = env.step(action)
+                ep_reward += reward
+                extrinsic_reward += reward
+                state = next_state
+                agent.update()
+        print(f'Episode:{i_ep+1}/{cfg.train_eps}, Reward:{ep_reward}, Loss:{agent.loss_numpy:.2f}, Meta_Loss:{agent.meta_loss_numpy:.2f}')
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(
+                0.9*ma_rewards[-1]+0.1*ep_reward)
+        else:
+            ma_rewards.append(ep_reward)
+    print('Complete training！')
+    return rewards, ma_rewards
+
+if __name__ == "__main__":
+    cfg = HierarchicalDQNConfig()
+
+    # train
+    env,agent = env_agent_config(cfg,seed=1)
+    rewards, ma_rewards = train(cfg, env, agent)
+    make_dir(cfg.result_path, cfg.model_path)
+    agent.save(path=cfg.model_path)
+    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)
+    save_results(rewards,ma_rewards,tag='eval',path=cfg.result_path)
+    plot_rewards(rewards,ma_rewards,tag="eval",env=cfg.env,algo = cfg.algo,path=cfg.result_path)
+

codes/DQN-series/NoisyDQN/task0_train.ipynb

@@ -0,0 +1,25 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import sys\n",
+    "from pathlib import Path\n",
+    "curr_path = str(Path().absolute()) # 当前路径\n",
+    "parent_path = str(Path().absolute().parent) # 父路径\n",
+    "sys.path.append(parent_path) # 添加路径到系统路径"
+   ]
+  }
+ ],
+ "metadata": {
+  "language_info": {
+   "name": "python"
+  },
+  "orig_nbformat": 4
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

codes/DQN-series/README.md

@@ -0,0 +1,3 @@
+
+本目录下汇总了基础的DQN及其变种或升级，如下
+