update

codes/DDPG/agent.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-09 20:25:52
 @LastEditor: John
-LastEditTime: 2021-03-17 20:43:25
+LastEditTime: 2021-03-31 00:56:32
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -58,9 +58,7 @@ class DDPG:
         done = torch.FloatTensor(np.float32(done)).unsqueeze(1).to(self.device)
         # 注意critic将(s_t,a)作为输入
         policy_loss = self.critic(state, self.actor(state))
-        
         policy_loss = -policy_loss.mean()
-
         next_action = self.target_actor(next_state)
         target_value = self.target_critic(next_state, next_action.detach())
         expected_value = reward + (1.0 - done) * self.gamma * target_value
@@ -87,7 +85,7 @@ class DDPG:
                 param.data * self.soft_tau
             )
     def save(self,path):
-        torch.save(self.target_net.state_dict(), path+'DDPG_checkpoint.pth')
+        torch.save(self.actor.state_dict(), path+'checkpoint.pt')
 
     def load(self,path):
-        self.actor.load_state_dict(torch.load(path+'DDPG_checkpoint.pth')) 
+        self.actor.load_state_dict(torch.load(path+'checkpoint.pt')) 

codes/DDPG/main.py

@@ -5,12 +5,17 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-11 20:58:21
 @LastEditor: John
-LastEditTime: 2021-03-19 19:57:00
+LastEditTime: 2021-03-31 01:04:48
 @Discription: 
 @Environment: python 3.7.7
 '''
 import sys,os
-sys.path.append(os.getcwd()) # 添加当前终端路径
+from pathlib import Path
+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 torch
 import gym
 import numpy as np
@@ -20,27 +25,23 @@ from DDPG.env import NormalizedActions,OUNoise
 from common.plot import plot_rewards
 from common.utils import save_results
 
-SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # 获取当前时间
-SAVED_MODEL_PATH = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"+SEQUENCE+'/' # 生成保存的模型路径
-if not os.path.exists(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"): # 检测是否存在文件夹
-    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/")
-if not os.path.exists(SAVED_MODEL_PATH): # 检测是否存在文件夹
-    os.mkdir(SAVED_MODEL_PATH)
-RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/results/"+SEQUENCE+'/' # 存储reward的路径
-if not os.path.exists(os.path.split(os.path.abspath(__file__))[0]+"/results/"): # 检测是否存在文件夹
-    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/results/")
-if not os.path.exists(RESULT_PATH): # 检测是否存在文件夹
-    os.mkdir(RESULT_PATH)
+SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
+SAVED_MODEL_PATH = curr_path+"/saved_model/"+SEQUENCE+'/' # path to save model
+if not os.path.exists(curr_path+"/saved_model/"): os.mkdir(curr_path+"/saved_model/")
+if not os.path.exists(SAVED_MODEL_PATH): os.mkdir(SAVED_MODEL_PATH)
+RESULT_PATH = curr_path+"/results/"+SEQUENCE+'/' # path to save rewards
+if not os.path.exists(curr_path+"/results/"): os.mkdir(curr_path+"/results/")
+if not os.path.exists(RESULT_PATH): os.mkdir(RESULT_PATH)
 
 class DDPGConfig:
     def __init__(self):
+        self.algo = 'DDPG'
         self.gamma = 0.99
         self.critic_lr = 1e-3  
         self.actor_lr = 1e-4 
         self.memory_capacity = 10000
         self.batch_size = 128
         self.train_eps =300
-        self.train_steps = 200
         self.eval_eps = 200
         self.eval_steps = 200
         self.target_update = 4
@@ -56,19 +57,19 @@ def train(cfg,env,agent):
     for i_episode in range(cfg.train_eps):
         state = env.reset()
         ou_noise.reset()
+        done = False
         ep_reward = 0
-        for i_step in range(cfg.train_steps):
+        i_step = 0
+        while not done:
+            i_step += 1
             action = agent.choose_action(state)
-            action = ou_noise.get_action(
-                action, i_step)  # 即paper中的random process
+            action = ou_noise.get_action(action, i_step)  # 即paper中的random process
             next_state, reward, done, _ = env.step(action)
             ep_reward += reward
             agent.memory.push(state, action, reward, next_state, done)
             agent.update()
             state = next_state
-            if done:
-                break
-        print('Episode:{}/{}, Reward:{}, Steps:{}, Done:{}'.format(i_episode+1,cfg.train_eps,ep_reward,i_step+1,done))
+        print('Episode:{}/{}, Reward:{}'.format(i_episode+1,cfg.train_eps,ep_reward))
         ep_steps.append(i_step)
         rewards.append(ep_reward)
         if ma_rewards:

codes/DQN/README.md

@@ -1,7 +1,7 @@
 # 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/leedeeprl-notes/#/chapter6/chapter6)。
+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。
 

codes/DQN/agent.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:50:49
 @LastEditor: John
-LastEditTime: 2021-03-13 14:56:23
+LastEditTime: 2021-03-30 17:01:26
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -13,6 +13,8 @@ LastEditTime: 2021-03-13 14:56:23
 '''
 
 
+
+
 import torch
 import torch.nn as nn
 import torch.optim as optim
@@ -23,61 +25,44 @@ from common.memory import ReplayBuffer
 from common.model import MLP
 class DQN:
     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 # 奖励的折扣因子
+        self.gamma = cfg.gamma  # 奖励的折扣因子
         # e-greedy策略相关参数
-        self.sample_count = 0 # 用于epsilon的衰减计数
-        self.epsilon = 0
-        self.epsilon_start = cfg.epsilon_start
-        self.epsilon_end = cfg.epsilon_end
-        self.epsilon_decay = cfg.epsilon_decay
+        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(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
+        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)
         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, train=True):
+    def choose_action(self, state):
         '''选择动作
         '''
-        if train:
-            self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \
-                math.exp(-1. * self.sample_count / self.epsilon_decay)
-            self.sample_count += 1
-            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.action_dim)
-            return action
-        else: 
-            with torch.no_grad(): # 取消保存梯度
-                    # 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='cpu', dtype=torch.float32) # 如tensor([[-0.0798, -0.0079]], grad_fn=<AddmmBackward>)
-                    q_value = self.target_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
+        self.frame_idx += 1
+        if random.random() > self.epsilon(self.frame_idx):
+            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.action_dim)
+        return action
+
     def update(self):
 
         if len(self.memory) < self.batch_size:
@@ -96,32 +81,31 @@ class DQN:
         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).unsqueeze(1)  # 将bool转为float然后转为张量
+            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_state_values = self.target_net(
-            next_state_batch).max(1)[0].detach()  # 比如tensor([ 0.0060, -0.0171,...,])
+        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_state_values * (1-done_batch[0])
+        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
         # 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)
-            
+        # 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):
+    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'))  
+    def load(self, path):
+        self.target_net.load_state_dict(torch.load(path+'dqn_checkpoint.pth'))

codes/DQN/main.py

@@ -5,12 +5,17 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:48:57
 @LastEditor: John
-LastEditTime: 2021-03-26 17:17:17
+LastEditTime: 2021-03-30 16:59:19
 @Discription: 
 @Environment: python 3.7.7
 '''
 import sys,os
-sys.path.append(os.getcwd()) # 添加当前终端路径
+from pathlib import Path
+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
@@ -18,58 +23,52 @@ from DQN.agent import DQN
 from common.plot import plot_rewards
 from common.utils import save_results
 
-SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # 获取当前时间
-SAVED_MODEL_PATH = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"+SEQUENCE+'/' # 生成保存的模型路径
-if not os.path.exists(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"): # 检测是否存在文件夹
-    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/")
-if not os.path.exists(SAVED_MODEL_PATH): # 检测是否存在文件夹
+SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
+SAVED_MODEL_PATH = curr_path+"/saved_model/"+SEQUENCE+'/' # path to save model
+if not os.path.exists(curr_path+"/saved_model/"): 
+    os.mkdir(curr_path+"/saved_model/")
+if not os.path.exists(SAVED_MODEL_PATH): 
     os.mkdir(SAVED_MODEL_PATH)
-RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/results/"+SEQUENCE+'/' # 存储reward的路径
-if not os.path.exists(os.path.split(os.path.abspath(__file__))[0]+"/results/"): # 检测是否存在文件夹
-    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/results/")
-if not os.path.exists(RESULT_PATH): # 检测是否存在文件夹
+RESULT_PATH = curr_path+"/results/"+SEQUENCE+'/' # path to save rewards
+if not os.path.exists(curr_path+"/results/"): 
+    os.mkdir(curr_path+"/results/")
+if not os.path.exists(RESULT_PATH): 
     os.mkdir(RESULT_PATH)
 
 class DQNConfig:
     def __init__(self):
-        self.algo = "DQN" # 算法名称
-        self.gamma = 0.99
-        self.epsilon_start = 0.95 # e-greedy策略的初始epsilon
+        self.algo = "DQN"  # name of algo
+        self.gamma = 0.95
+        self.epsilon_start = 1 # e-greedy策略的初始epsilon
         self.epsilon_end = 0.01
-        self.epsilon_decay = 200
-        self.lr = 0.01 # 学习率
-        self.memory_capacity = 800 # Replay Memory容量
-        self.batch_size = 64
+        self.epsilon_decay = 500
+        self.lr = 0.0001 # learning rate
+        self.memory_capacity = 10000 # Replay Memory容量
+        self.batch_size = 32
         self.train_eps = 300 # 训练的episode数目
-        self.train_steps = 200 # 训练每个episode的最大长度
         self.target_update = 2 # target net的更新频率
         self.eval_eps = 20 # 测试的episode数目
-        self.eval_steps = 200 # 测试每个episode的最大长度
         self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
-        self.hidden_dim = 128 # 神经网络隐藏层维度
+        self.hidden_dim = 256 # 神经网络隐藏层维度
  
 def train(cfg,env,agent):
     print('Start to train !')
     rewards = []
-    ma_rewards = [] # 滑动平均的reward
-    ep_steps = []
+    ma_rewards = [] # moveing average reward
     for i_episode in range(cfg.train_eps):
-        state = env.reset() # reset环境状态
+        state = env.reset() 
+        done = False
         ep_reward = 0
-        for i_step in range(cfg.train_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
+        while not done:
+            action = agent.choose_action(state) 
+            next_state, reward, done, _ = env.step(action) 
             ep_reward += reward
-            agent.memory.push(state, action, reward, next_state, done) # 将state等这些transition存入memory
-            state = next_state # 跳转到下一个状态
-            agent.update() # 每步更新网络
-            if done:
-                break
-        # 更新target network，复制DQN中的所有weights and biases
+            agent.memory.push(state, action, reward, next_state, done) 
+            state = next_state 
+            agent.update() 
         if i_episode % cfg.target_update == 0:
             agent.target_net.load_state_dict(agent.policy_net.state_dict())
-        print('Episode:{}/{}, Reward:{}, Steps:{}, Done:{}'.format(i_episode+1,cfg.train_eps,ep_reward,i_step+1,done))
-        ep_steps.append(i_step)
+        print('Episode:{}/{}, Reward:{}'.format(i_episode+1,cfg.train_eps,ep_reward))
         rewards.append(ep_reward)
         # 计算滑动窗口的reward
         if ma_rewards:
@@ -82,8 +81,8 @@ def train(cfg,env,agent):
 
 if __name__ == "__main__":
     cfg = DQNConfig()
-    env = gym.make('CartPole-v0').unwrapped # 可google为什么unwrapped gym，此处一般不需要
-    env.seed(1) # 设置env随机种子
+    env = gym.make('CartPole-v0')
+    env.seed(1)
     state_dim = env.observation_space.shape[0]
     action_dim = env.action_space.n
     agent = DQN(state_dim,action_dim,cfg)

codes/DQN_cnn/main.py

@@ -5,12 +5,17 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-11 10:01:09
 @LastEditor: John
-LastEditTime: 2021-03-23 20:43:28
+LastEditTime: 2021-03-29 20:23:48
 @Discription: 
 @Environment: python 3.7.7
 '''
 import sys,os
-sys.path.append(os.getcwd()) # add current terminal path to sys.path
+from pathlib import Path
+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
@@ -19,17 +24,15 @@ from DQN_cnn.agent import DQNcnn
 from common.plot import plot_rewards
 from common.utils import save_results
 
-sys.path.append(os.getcwd())  # add current terminal path to sys.path
-
 SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
-SAVED_MODEL_PATH = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"+SEQUENCE+'/' # path to save model
-if not os.path.exists(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"): 
-    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/")
+SAVED_MODEL_PATH = curr_path+"/saved_model/"+SEQUENCE+'/' # path to save model
+if not os.path.exists(curr_path+"/saved_model/"): 
+    os.mkdir(curr_path+"/saved_model/")
 if not os.path.exists(SAVED_MODEL_PATH): 
     os.mkdir(SAVED_MODEL_PATH)
-RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/results/"+SEQUENCE+'/' # path to save rewards
-if not os.path.exists(os.path.split(os.path.abspath(__file__))[0]+"/results/"): 
-    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/results/")
+RESULT_PATH = curr_path+"/results/"+SEQUENCE+'/' # path to save rewards
+if not os.path.exists(curr_path+"/results/"): 
+    os.mkdir(curr_path+"/results/")
 if not os.path.exists(RESULT_PATH): 
     os.mkdir(RESULT_PATH)
 

codes/DoubleDQN/memory.py

@@ -1,40 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-@Author: John
-@Email: johnjim0816@gmail.com
-@Date: 2020-06-10 15:27:16
-@LastEditor: John
-LastEditTime: 2021-01-20 18:58:37
-@Discription: 
-@Environment: python 3.7.7
-'''
-import random
-
-class ReplayBuffer:
-    
-    def __init__(self, capacity):
-        self.capacity = capacity # buffer的最大容量
-        self.buffer = []
-        self.position = 0
-    
-    def push(self, state, action, reward, next_state, done):
-        '''以队列的方式将样本填入buffer中
-        '''
-        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):
-        '''随机采样batch_size个样本
-        '''
-        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):
-        '''返回buffer的长度
-        '''
-        return len(self.buffer)
-

codes/DoubleDQN/model.py

@@ -1,30 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-@Author: John
-@Email: johnjim0816@gmail.com
-@Date: 2020-06-12 00:47:02
-@LastEditor: John
-LastEditTime: 2020-08-19 16:55:54
-@Discription: 
-@Environment: python 3.7.7
-'''
-import torch.nn as nn
-import torch.nn.functional as F
-
-class MLP(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(MLP, 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)

codes/DoubleDQN/params.py

@@ -1,51 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-Author: John
-Email: johnjim0816@gmail.com
-Date: 2020-12-22 15:22:17
-LastEditor: John
-LastEditTime: 2021-01-21 14:30:38
-Discription: 
-Environment: 
-'''
-import datetime
-import os
-import argparse
-
-ALGO_NAME = 'Double DQN'
-SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
-SAVED_MODEL_PATH = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"+SEQUENCE+'/'
-RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/results/"+SEQUENCE+'/'
-
-TRAIN_LOG_DIR=os.path.split(os.path.abspath(__file__))[0]+"/logs/train/" + SEQUENCE
-EVAL_LOG_DIR=os.path.split(os.path.abspath(__file__))[0]+"/logs/eval/" + SEQUENCE
-
-def get_args():
-    '''模型参数
-    '''
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--train", default=1, type=int)  # 1 表示训练，0表示只进行eval
-    parser.add_argument("--gamma", default=0.99,
-                        type=float)  # q-learning中的gamma
-    parser.add_argument("--epsilon_start", default=0.95,
-                        type=float)  # 基于贪心选择action对应的参数epsilon
-    parser.add_argument("--epsilon_end", default=0.01, type=float)
-    parser.add_argument("--epsilon_decay", default=500, type=float)
-    parser.add_argument("--policy_lr", default=0.01, type=float)
-    parser.add_argument("--memory_capacity", default=1000,
-                        type=int, help="capacity of Replay Memory") 
-
-    parser.add_argument("--batch_size", default=32, type=int,
-                        help="batch size of memory sampling")
-    parser.add_argument("--train_eps", default=200, type=int) # 训练的最大episode数目
-    parser.add_argument("--train_steps", default=200, type=int)
-    parser.add_argument("--target_update", default=2, type=int,
-                        help="when(every default 2 eisodes) to update target net ") # 更新频率
-
-    parser.add_argument("--eval_eps", default=100, type=int)  # 训练的最大episode数目
-    parser.add_argument("--eval_steps", default=200,
-                        type=int)  # 训练每个episode的长度
-    config = parser.parse_args()
-
-    return config

codes/DoubleDQN/plot.py

@@ -1,48 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-@Author: John
-@Email: johnjim0816@gmail.com
-@Date: 2020-06-11 16:30:09
-@LastEditor: John
-LastEditTime: 2020-12-22 15:24:31
-@Discription: 
-@Environment: python 3.7.7
-'''
-import matplotlib.pyplot as plt
-import seaborn as sns
-import numpy as np
-import os 
-from params import ALGO_NAME
-def plot(item,ylabel='rewards_train', save_fig = True):
-    '''plot using searborn to plot 
-    '''
-    sns.set()
-    plt.figure()
-    plt.plot(np.arange(len(item)), item)
-    plt.title(ylabel+' of '+ALGO_NAME) 
-    plt.ylabel(ylabel)
-    plt.xlabel('episodes')
-    if save_fig:
-        plt.savefig(os.path.dirname(__file__)+"/results/"+ylabel+".png")
-    plt.show()
-
-
-    # plt.show()
-if __name__ == "__main__":
-
-    output_path = os.path.split(os.path.abspath(__file__))[0]+"/results/"
-    tag = 'train'
-    rewards=np.load(output_path+"rewards_"+tag+".npy", )
-    moving_average_rewards=np.load(output_path+"moving_average_rewards_"+tag+".npy",)
-    steps=np.load(output_path+"steps_"+tag+".npy")
-    plot(rewards)
-    plot(moving_average_rewards,ylabel='moving_average_rewards_'+tag)
-    plot(steps,ylabel='steps_'+tag)
-    tag = 'eval'
-    rewards=np.load(output_path+"rewards_"+tag+".npy", )
-    moving_average_rewards=np.load(output_path+"moving_average_rewards_"+tag+".npy",)
-    steps=np.load(output_path+"steps_"+tag+".npy")
-    plot(rewards,ylabel='rewards_'+tag)
-    plot(moving_average_rewards,ylabel='moving_average_rewards_'+tag)
-    plot(steps,ylabel='steps_'+tag) 

codes/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/HierarchicalDQN/agent.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-24 22:18:18
 LastEditor: John
-LastEditTime: 2021-03-27 04:24:30
+LastEditTime: 2021-03-31 14:51:09
 Discription: 
 Environment: 
 '''
@@ -13,90 +13,103 @@ import torch
 import torch.nn as nn
 import numpy as np
 import random,math
-from HierarchicalDQN.model import MLP
-from common.memory import ReplayBuffer
 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.sample_count = 0 
-        self.epsilon = 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.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.target_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.meta_target_net = MLP(state_dim, state_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)
-    def to_onehot(x):
-        oh = np.zeros(6)
+        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,meta_state):
-        self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * math.exp(-1. * self.sample_count / self.epsilon_decay)
-        self.sample_count += 1
-        if random.random() > self.epsilon:
+    def set_goal(self,state):
+        if random.random() > self.epsilon(self.frame_idx):
             with torch.no_grad():
-                meta_state = torch.tensor([meta_state], device=self.device, dtype=torch.float32)
-                q_value = self.policy_net(meta_state)
-                goal = q_value.max(1)[1].item() 
+                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.action_dim)
-        goal = self.meta_policy_net(meta_state)
-        onehot_goal = self.to_onehot(goal)
-        return onehot_goal
+            goal = random.randrange(self.state_dim)
+        return goal
     def choose_action(self,state):
-        self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * math.exp(-1. * self.sample_count / self.epsilon_decay)
-        self.sample_count += 1
-        if random.random() > self.epsilon:
+        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)
+                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):
-            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).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).unsqueeze(1)  
-        q_values = self.policy_net(state_batch).gather(dim=1, index=action_batch)
-        next_state_values = self.target_net(next_state_batch).max(1)[0].detach()  
-        expected_q_values = reward_batch + self.gamma * next_state_values * (1-done_batch[0])
-        loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1)) 
+            return
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(self.batch_size)
+        state_batch = torch.tensor(state_batch,dtype=torch.float)
+        action_batch = torch.tensor(action_batch,dtype=torch.int64).unsqueeze(1)  
+        reward_batch = torch.tensor(reward_batch,dtype=torch.float)  
+        next_state_batch = torch.tensor(next_state_batch, dtype=torch.float)
+        done_batch = torch.tensor(np.float32(done_batch))
+        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(): 
+        for param in self.policy_net.parameters():  # clip防止梯度爆炸
             param.grad.data.clamp_(-1, 1)
-        self.optimizer.step() 
-
+        self.optimizer.step()  
+        self.loss_numpy = loss.detach().numpy()
+        self.losses.append(self.loss_numpy)  
+    def update_meta(self):
         if self.batch_size > len(self.meta_memory):
-            meta_state_batch, meta_action_batch, meta_reward_batch, next_meta_state_batch, meta_done_batch = self.memory.sample(self.batch_size)
-        meta_state_batch = torch.tensor(meta_state_batch, device=self.device, dtype=torch.float)
-        meta_action_batch = torch.tensor(meta_action_batch, device=self.device).unsqueeze(1)  
-        meta_reward_batch = torch.tensor(meta_reward_batch, device=self.device, dtype=torch.float)  
-        next_meta_state_batch = torch.tensor(next_meta_state_batch, device=self.device, dtype=torch.float)
-        meta_done_batch = torch.tensor(np.float32(meta_done_batch), device=self.device).unsqueeze(1)  
-        meta_q_values = self.meta_policy_net(meta_state_batch).gather(dim=1, index=meta_action_batch)
-        next_state_values = self.target_net(next_meta_state_batch).max(1)[0].detach()  
-        expected_meta_q_values = meta_reward_batch + self.gamma * next_state_values * (1-meta_done_batch[0])
-        meta_loss = nn.MSEmeta_loss()(meta_q_values, expected_meta_q_values.unsqueeze(1)) 
+            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,dtype=torch.float)
+        action_batch = torch.tensor(action_batch,dtype=torch.int64).unsqueeze(1)  
+        reward_batch = torch.tensor(reward_batch,dtype=torch.float)  
+        next_state_batch = torch.tensor(next_state_batch, dtype=torch.float)
+        done_batch = torch.tensor(np.float32(done_batch))
+        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(): 
+        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().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/HierarchicalDQN/main.py

@@ -3,95 +3,108 @@
 '''
 Author: John
 Email: johnjim0816@gmail.com
-Date: 2021-03-24 22:14:04
+Date: 2021-03-29 10:37:32
 LastEditor: John
-LastEditTime: 2021-03-27 04:23:43
+LastEditTime: 2021-03-31 14:58:49
 Discription: 
 Environment: 
 '''
+
+
 import sys,os
-sys.path.append(os.getcwd()) # add current terminal path to sys.path
-import gym
+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 datetime
-from HierarchicalDQN.agent import HierarchicalDQN
-from common.plot import plot_rewards
-from common.utils import save_results
+import gym
 
-SEQUENCE = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
-SAVED_MODEL_PATH = os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"+SEQUENCE+'/'  # path to save model
-if not os.path.exists(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/"): 
-    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/saved_model/")
+from common.utils import save_results
+from common.plot import plot_rewards,plot_losses
+from HierarchicalDQN.agent import HierarchicalDQN
+
+SEQUENCE = datetime.datetime.now().strftime(
+    "%Y%m%d-%H%M%S")  # obtain current time
+SAVED_MODEL_PATH = curr_path+"/saved_model/"+SEQUENCE+'/'  # path to save model
+if not os.path.exists(curr_path+"/saved_model/"):
+    os.mkdir(curr_path+"/saved_model/")
 if not os.path.exists(SAVED_MODEL_PATH):
     os.mkdir(SAVED_MODEL_PATH)
-RESULT_PATH = os.path.split(os.path.abspath(__file__))[0]+"/results/"+SEQUENCE+'/' # path to save rewards
-if not os.path.exists(os.path.split(os.path.abspath(__file__))[0]+"/results/"): 
-    os.mkdir(os.path.split(os.path.abspath(__file__))[0]+"/results/")
-if not os.path.exists(RESULT_PATH): 
+RESULT_PATH = curr_path+"/results/"+SEQUENCE+'/'  # path to save rewards
+if not os.path.exists(curr_path+"/results/"):
+    os.mkdir(curr_path+"/results/")
+if not os.path.exists(RESULT_PATH):
     os.mkdir(RESULT_PATH)
 
+
 class HierarchicalDQNConfig:
     def __init__(self):
-        self.algo = "DQN" # name of algo
+        self.algo = "H-DQN"  # name of algo
         self.gamma = 0.99
-        self.epsilon_start = 0.95 # start epsilon of e-greedy policy
+        self.epsilon_start = 1  # start epsilon of e-greedy policy
         self.epsilon_end = 0.01
         self.epsilon_decay = 200
-        self.lr = 0.01 # learning rate
-        self.memory_capacity = 800 # Replay Memory capacity
-        self.batch_size = 64
-        self.train_eps = 250 # 训练的episode数目
-        self.train_steps = 200 # 训练每个episode的最大长度
-        self.target_update = 2 # target net的更新频率
-        self.eval_eps = 20 # 测试的episode数目
-        self.eval_steps = 200 # 测试每个episode的最大长度
-        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测gpu
-        self.hidden_dim = 256 # dimension of hidden layer
+        self.lr = 0.0001  # learning rate
+        self.memory_capacity = 10000  # Replay Memory capacity
+        self.batch_size = 32
+        self.train_eps = 300  # 训练的episode数目
+        self.target_update = 2  # target net的更新频率
+        self.eval_eps = 20  # 测试的episode数目
+        self.device = torch.device(
+            "cuda" if torch.cuda.is_available() else "cpu")  # 检测gpu
+        self.hidden_dim = 256  # dimension of hidden layer
 
-def train(cfg,env,agent):
+
+def train(cfg, env, agent):
     print('Start to train !')
     rewards = []
-    ma_rewards = [] # moving average reward
-    ep_steps = []
+    ma_rewards = []  # moveing average reward
     for i_episode in range(cfg.train_eps):
-        state = env.reset() 
-        extrinsic_reward = 0
-        for i_step in range(cfg.train_steps):
-            goal= agent.set_goal(state)
+        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
-            goal_state  = np.concatenate([state, goal])
-            action = agent.choose_action(state) 
-            next_state, reward, done, _ = env.step(action)
-            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, goal]), done)
-            state = next_state 
-            agent.update()
-            if done:
-                break
-        if i_episode % cfg.target_update == 0:
-            agent.target_net.load_state_dict(agent.policy_net.state_dict())
-        print('Episode:{}/{}, Reward:{}, Steps:{}, Done:{}'.format(i_episode+1,cfg.train_eps,extrinsic_reward,i_step+1,done))
-        ep_steps.append(i_step)
-        rewards.append(extrinsic_reward)
+            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_episode+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*extrinsic_reward)
+                0.9*ma_rewards[-1]+0.1*ep_reward)
         else:
-            ma_rewards.append(extrinsic_reward)   
-    agent.meta_memory.push(meta_state, goal, extrinsic_reward, state, done)
+            ma_rewards.append(ep_reward)
     print('Complete training！')
-    return rewards,ma_rewards
+    return rewards, ma_rewards
+
 
 if __name__ == "__main__":
-    cfg = HierarchicalDQNConfig()
     env = gym.make('CartPole-v0')
-    env.seed(1) 
+    env.seed(1)
+    cfg = HierarchicalDQNConfig()
     state_dim = env.observation_space.shape[0]
     action_dim = env.action_space.n
-    agent = HierarchicalDQN(state_dim,action_dim,cfg)
-    rewards,ma_rewards = train(cfg,env,agent)
+    agent = HierarchicalDQN(state_dim, action_dim, cfg)
+    rewards, ma_rewards = train(cfg, env, agent)
     agent.save(path=SAVED_MODEL_PATH)
-    save_results(rewards,ma_rewards,tag='train',path=RESULT_PATH)
-    plot_rewards(rewards,ma_rewards,tag="train",algo = cfg.algo,path=RESULT_PATH)
+    save_results(rewards, ma_rewards, tag='train', path=RESULT_PATH)
+    plot_rewards(rewards, ma_rewards, tag="train",
+                 algo=cfg.algo, path=RESULT_PATH)
+    plot_losses(agent.losses,algo=cfg.algo, path=RESULT_PATH)
+

codes/HierarchicalDQN/model.py

@@ -1,24 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-Author: John
-Email: johnjim0816@gmail.com
-Date: 2021-03-24 22:14:12
-LastEditor: John
-LastEditTime: 2021-03-24 22:17:09
-Discription: 
-Environment: 
-'''
-import torch.nn as nn
-import torch.nn.functional as F
-class MLP(nn.Module):
-    def __init__(self, state_dim,action_dim,hidden_dim=128):
-        super(MLP, self).__init__()
-        self.fc1 = nn.Linear(state_dim, hidden_dim) 
-        self.fc2 = nn.Linear(hidden_dim,hidden_dim)
-        self.fc3 = nn.Linear(hidden_dim, action_dim) 
-        
-    def forward(self, x):
-        x = F.relu(self.fc1(x)) 
-        x = F.relu(self.fc2(x))
-        return self.fc3(x)

codes/README.md

@@ -19,9 +19,10 @@
 ## 运行环境
 
 python 3.7、pytorch 1.6.0-1.7.1、gym 0.17.0-0.18.0
+
 ## 使用说明
 
-对应算法文件夹下运行```main.py```即可
+运行```main.py```或者```main.ipynb```
 ## 算法进度
 
 |                 算法名称                 |                        相关论文材料                         | 环境                                  |                备注                |
@@ -29,17 +30,17 @@ python 3.7、pytorch 1.6.0-1.7.1、gym 0.17.0-0.18.0
 | [On-Policy First-Visit MC](./MonteCarlo) |                                                             | [Racetrack](./envs/racetrack_env.md)  |                                    |
 |        [Q-Learning](./QLearning)         |                                                             | [CliffWalking-v0](./envs/gym_info.md) |                                    |
 |             [Sarsa](./Sarsa)             |                                                             | [Racetrack](./envs/racetrack_env.md)  |                                    |
-|               [DQN](./DQN)               | [DQN-paper](https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf) | [CartPole-v0](./envs/gym_info.md)     |                                    |
-|                 [DQN-cnn](./DQN_cnn)                  | [DQN-paper](https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf) | [CartPole-v0](./envs/gym_info.md)     | 与DQN相比使用了CNN而不是全链接网络 |
+|               [DQN](./DQN)               | [DQN Paper](https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf) | [CartPole-v0](./envs/gym_info.md)     |                                    |
+|           [DQN-cnn](./DQN_cnn)           | [DQN Paper](https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf) | [CartPole-v0](./envs/gym_info.md)     | 与DQN相比使用了CNN而不是全链接网络 |
 |         [DoubleDQN](./DoubleDQN)         |                                                             | [CartPole-v0](./envs/gym_info.md)     |          效果不好，待改进          |
-|             Hierarchical DQN             |    [Hierarchical DQN](https://arxiv.org/abs/1604.06057)     |                                       |                                    |
+|             Hierarchical DQN             |       [H-DQN Paper](https://arxiv.org/abs/1604.06057)       |                                       |                                    |
 |    [PolicyGradient](./PolicyGradient)    |                                                             | [CartPole-v0](./envs/gym_info.md)     |                                    |
 |                   A2C                    |                                                             | [CartPole-v0](./envs/gym_info.md)     |                                    |
 |                   A3C                    |                                                             |                                       |                                    |
 |                   SAC                    |                                                             |                                       |                                    |
 |               [PPO](./PPO)               |        [PPO paper](https://arxiv.org/abs/1707.06347)        | [CartPole-v0](./envs/gym_info.md)     |                                    |
 |                   DDPG                   |       [DDPG Paper](https://arxiv.org/abs/1509.02971)        | [Pendulum-v0](./envs/gym_info.md)     |                                    |
-|                   TD3                    | [Twin Dueling DDPG Paper](https://arxiv.org/abs/1802.09477) |                                       |                                    |
+|                   TD3                    |        [TD3 Paper](https://arxiv.org/abs/1802.09477)        |                                       |                                    |
 |                   GAIL                   |                                                             |                                       |                                    |
 
 

codes/README_en.md

@@ -24,7 +24,7 @@ Note that ```model.py```,```memory.py```,```plot.py``` shall be utilized in diff
 python 3.7.9、pytorch 1.6.0、gym 0.18.0
 ## Usage
 
-Environment infomations see [环境说明](https://github.com/JohnJim0816/reinforcement-learning-tutorials/blob/master/env_info.md)
+run ```main.py``` or ```main.ipynb```
 
 ## Schedule
 

codes/common/model.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-12 21:14:12
 LastEditor: John
-LastEditTime: 2021-03-24 22:15:00
+LastEditTime: 2021-03-31 13:49:06
 Discription: 
 Environment: 
 '''
@@ -15,15 +15,15 @@ import torch.nn.functional as F
 from torch.distributions import Categorical
 
 class MLP(nn.Module):
-    def __init__(self, state_dim,action_dim,hidden_dim=128):
+    def __init__(self, input_dim,output_dim,hidden_dim=128):
         """ 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
+            input_dim: 输入的feature即环境的state数目
+            output_dim: 输出的action总个数
         """
         super(MLP, self).__init__()
-        self.fc1 = nn.Linear(state_dim, hidden_dim) # 输入层
+        self.fc1 = nn.Linear(input_dim, hidden_dim) # 输入层
         self.fc2 = nn.Linear(hidden_dim,hidden_dim) # 隐藏层
-        self.fc3 = nn.Linear(hidden_dim, action_dim) # 输出层
+        self.fc3 = nn.Linear(hidden_dim, output_dim) # 输出层
         
     def forward(self, x):
         # 各层对应的激活函数
@@ -32,10 +32,10 @@ class MLP(nn.Module):
         return self.fc3(x)
 
 class Critic(nn.Module):
-    def __init__(self, n_obs, action_dim, hidden_size, init_w=3e-3):
+    def __init__(self, n_obs, output_dim, hidden_size, init_w=3e-3):
         super(Critic, self).__init__()
         
-        self.linear1 = nn.Linear(n_obs + action_dim, hidden_size)
+        self.linear1 = nn.Linear(n_obs + output_dim, hidden_size)
         self.linear2 = nn.Linear(hidden_size, hidden_size)
         self.linear3 = nn.Linear(hidden_size, 1)
         # 随机初始化为较小的值
@@ -51,11 +51,11 @@ class Critic(nn.Module):
         return x
 
 class Actor(nn.Module):
-    def __init__(self, n_obs, action_dim, hidden_size, init_w=3e-3):
+    def __init__(self, n_obs, output_dim, hidden_size, init_w=3e-3):
         super(Actor, self).__init__()  
         self.linear1 = nn.Linear(n_obs, hidden_size)
         self.linear2 = nn.Linear(hidden_size, hidden_size)
-        self.linear3 = nn.Linear(hidden_size, action_dim)
+        self.linear3 = nn.Linear(hidden_size, output_dim)
         
         self.linear3.weight.data.uniform_(-init_w, init_w)
         self.linear3.bias.data.uniform_(-init_w, init_w)
@@ -67,18 +67,18 @@ class Actor(nn.Module):
         return x
 
 class ActorCritic(nn.Module):
-    def __init__(self, state_dim, action_dim, hidden_dim=256):
+    def __init__(self, input_dim, output_dim, hidden_dim=256):
         super(ActorCritic, self).__init__()
         self.critic = nn.Sequential(
-            nn.Linear(state_dim, hidden_dim),
+            nn.Linear(input_dim, hidden_dim),
             nn.ReLU(),
             nn.Linear(hidden_dim, 1)
         )
         
         self.actor = nn.Sequential(
-            nn.Linear(state_dim, hidden_dim),
+            nn.Linear(input_dim, hidden_dim),
             nn.ReLU(),
-            nn.Linear(hidden_dim, action_dim),
+            nn.Linear(hidden_dim, output_dim),
             nn.Softmax(dim=1),
         )
         

codes/common/plot.py

@@ -5,13 +5,13 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-10-07 20:57:11
 LastEditor: John
-LastEditTime: 2021-03-13 11:31:49
+LastEditTime: 2021-03-31 14:05:52
 Discription: 
 Environment: 
 '''
 import matplotlib.pyplot as plt
 import seaborn as sns
-def plot_rewards(rewards,ma_rewards,tag="train",algo = "On-Policy First-Visit MC Control",path='./'):
+def plot_rewards(rewards,ma_rewards,tag="train",algo = "DQN",path='./'):
     sns.set()
     plt.title("average learning curve of {}".format(algo))
     plt.xlabel('epsiodes')
@@ -20,4 +20,13 @@ def plot_rewards(rewards,ma_rewards,tag="train",algo = "On-Policy First-Visit MC
     plt.legend()
     plt.savefig(path+"rewards_curve_{}".format(tag))
     plt.show()
+
+def plot_losses(losses,algo = "DQN",path='./'):
+    sns.set()
+    plt.title("loss curve of {}".format(algo))
+    plt.xlabel('epsiodes')
+    plt.plot(losses,label='rewards')
+    plt.legend()
+    plt.savefig(path+"losses_curve")
+    plt.show()