update projects

projects/codes/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。

projects/codes/DoubleDQN/double_dqn.py

@@ -0,0 +1,160 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+@Author: John
+@Email: johnjim0816@gmail.com
+@Date: 2020-06-12 00:50:49
+@LastEditor: John
+LastEditTime: 2022-07-21 00:08:26
+@Discription: 
+@Environment: python 3.7.7
+'''
+'''off-policy
+'''
+
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import torch.nn.functional as F
+import random
+import math
+import numpy as np
+
+class ReplayBuffer:
+    def __init__(self, capacity):
+        self.capacity = capacity # 经验回放的容量
+        self.buffer = [] # 缓冲区
+        self.position = 0 
+    
+    def push(self, state, action, reward, next_state, done):
+        ''' 缓冲区是一个队列，容量超出时去掉开始存入的转移(transition)
+        '''
+        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) # 随机采出小批量转移
+        state, action, reward, next_state, done =  zip(*batch) # 解压成状态，动作等
+        return state, action, reward, next_state, done
+    
+    def __len__(self):
+        ''' 返回当前存储的量
+        '''
+        return len(self.buffer)
+
+class MLP(nn.Module):
+    def __init__(self, n_states,n_actions,hidden_dim=128):
+        """ 初始化q网络，为全连接网络
+            n_states: 输入的特征数即环境的状态维度
+            n_actions: 输出的动作维度
+        """
+        super(MLP, self).__init__()
+        self.fc1 = nn.Linear(n_states, hidden_dim) # 输入层
+        self.fc2 = nn.Linear(hidden_dim,hidden_dim) # 隐藏层
+        self.fc3 = nn.Linear(hidden_dim, 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)
+        
+class DoubleDQN:
+    def __init__(self, n_states, n_actions, cfg):
+        self.n_actions = n_actions  # 总的动作个数
+        self.device = torch.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(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)
+        # 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 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:
+            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()  
+        else:
+            action = random.randrange(self.n_actions)
+        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)  

projects/codes/DoubleDQN/outputs/CartPole-v0/20220721-215416/results/params.json

@@ -0,0 +1,19 @@
+{
+    "algo_name": "DoubleDQN",
+    "env_name": "CartPole-v0",
+    "train_eps": 200,
+    "test_eps": 20,
+    "gamma": 0.99,
+    "epsilon_start": 0.95,
+    "epsilon_end": 0.01,
+    "epsilon_decay": 500,
+    "lr": 0.0001,
+    "memory_capacity": 100000,
+    "batch_size": 64,
+    "target_update": 2,
+    "hidden_dim": 256,
+    "device": "cuda",
+    "result_path": "C:\\Users\\24438\\Desktop\\rl-tutorials\\codes\\DoubleDQN/outputs/CartPole-v0/20220721-215416/results/",
+    "model_path": "C:\\Users\\24438\\Desktop\\rl-tutorials\\codes\\DoubleDQN/outputs/CartPole-v0/20220721-215416/models/",
+    "save_fig": true
+}

projects/codes/DoubleDQN/task0.py

@@ -0,0 +1,138 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: JiangJi
+Email: johnjim0816@gmail.com
+Date: 2021-11-07 18:10:37
+LastEditor: JiangJi
+LastEditTime: 2022-07-21 21:52:31
+Discription: 
+'''
+import sys,os
+curr_path = os.path.dirname(os.path.abspath(__file__))  # current path
+parent_path = os.path.dirname(curr_path)  # parent path
+sys.path.append(parent_path)  # add to system path
+
+import gym
+import torch
+import datetime
+import argparse
+
+from common.utils import save_results,make_dir
+from common.utils import plot_rewards,save_args
+from DoubleDQN.double_dqn import DoubleDQN
+
+def get_args():
+    """ Hyperparameters
+    """
+    curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")  # Obtain current time
+    parser = argparse.ArgumentParser(description="hyperparameters")      
+    parser.add_argument('--algo_name',default='DoubleDQN',type=str,help="name of algorithm")
+    parser.add_argument('--env_name',default='CartPole-v0',type=str,help="name of environment")
+    parser.add_argument('--train_eps',default=200,type=int,help="episodes of training")
+    parser.add_argument('--test_eps',default=20,type=int,help="episodes of testing")
+    parser.add_argument('--gamma',default=0.99,type=float,help="discounted factor")
+    parser.add_argument('--epsilon_start',default=0.95,type=float,help="initial value of epsilon")
+    parser.add_argument('--epsilon_end',default=0.01,type=float,help="final value of epsilon")
+    parser.add_argument('--epsilon_decay',default=500,type=int,help="decay rate of epsilon")
+    parser.add_argument('--lr',default=0.0001,type=float,help="learning rate")
+    parser.add_argument('--memory_capacity',default=100000,type=int,help="memory capacity")
+    parser.add_argument('--batch_size',default=64,type=int)
+    parser.add_argument('--target_update',default=2,type=int)
+    parser.add_argument('--hidden_dim',default=256,type=int)
+    parser.add_argument('--device',default='cpu',type=str,help="cpu or cuda") 
+    parser.add_argument('--result_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/results/' )
+    parser.add_argument('--model_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/models/' ) # path to save models
+    parser.add_argument('--save_fig',default=True,type=bool,help="if save figure or not")           
+    args = parser.parse_args()                          
+    return args
+        
+        
+def env_agent_config(cfg,seed=1):
+    env = gym.make(cfg.env_name)  
+    env.seed(seed)
+    n_states = env.observation_space.shape[0]
+    n_actions = env.action_space.n
+    agent = DoubleDQN(n_states,n_actions,cfg)
+    return env,agent
+
+def train(cfg,env,agent):
+    print('Start training!')
+    print(f'Env:{cfg.env_name}, Algorithm:{cfg.algo_name}, Device:{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) 
+            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())
+        if (i_ep+1)%10 == 0: 
+            print(f'Env:{i_ep+1}/{cfg.train_eps}, Reward:{ep_reward:.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('Finish training!')
+    return {'rewards':rewards,'ma_rewards':ma_rewards}
+
+def test(cfg,env,agent):
+    print('Start testing')
+    print(f'Env:{cfg.env_name}, Algorithm:{cfg.algo_name}, Device:{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.test_eps):
+        state = env.reset() 
+        ep_reward = 0   
+        while True:
+            action = agent.choose_action(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"Epside:{i_ep+1}/{cfg.test_eps}, Reward:{ep_reward:.1f}")
+    print('Finish testing!')
+    return {'rewards':rewards,'ma_rewards':ma_rewards}
+
+if __name__ == "__main__":
+    cfg = get_args()
+    print(cfg.device)
+    # training
+    env,agent = env_agent_config(cfg,seed=1)
+    res_dic = train(cfg, env, agent)
+    make_dir(cfg.result_path, cfg.model_path)
+    save_args(cfg)
+    agent.save(path=cfg.model_path)
+    save_results(res_dic, tag='train',
+                 path=cfg.result_path)  
+    plot_rewards(res_dic['rewards'], res_dic['ma_rewards'], cfg, tag="train")  
+    # testing
+    env,agent = env_agent_config(cfg,seed=10)
+    agent.load(path=cfg.model_path)
+    res_dic = test(cfg,env,agent)
+    save_results(res_dic, tag='test',
+                 path=cfg.result_path)  
+    plot_rewards(res_dic['rewards'], res_dic['ma_rewards'], cfg, tag="test")