update

codes/A2C/env.py

@@ -14,16 +14,16 @@ import gym
 from A2C.multiprocessing_env import SubprocVecEnv
 
 # num_envs = 16
-# env_name = "Pendulum-v0"
+# env = "Pendulum-v0"
 
-def make_envs(num_envs=16,env_name="Pendulum-v0"):
+def make_envs(num_envs=16,env="Pendulum-v0"):
     ''' 创建多个子环境
     '''
     num_envs = 16
-    env_name = "CartPole-v0"
+    env = "CartPole-v0"
     def make_env():
         def _thunk():
-            env = gym.make(env_name)
+            env = gym.make(env)
             return env
 
         return _thunk
@@ -34,10 +34,10 @@ def make_envs(num_envs=16,env_name="Pendulum-v0"):
 # if __name__ == "__main__":
 
 #     num_envs = 16
-#     env_name = "CartPole-v0"
+#     env = "CartPole-v0"
 #     def make_env():
 #         def _thunk():
-#             env = gym.make(env_name)
+#             env = gym.make(env)
 #             return env
 
 #         return _thunk
@@ -45,4 +45,4 @@ def make_envs(num_envs=16,env_name="Pendulum-v0"):
 #     envs = [make_env() for i in range(num_envs)]
 #     envs = SubprocVecEnv(envs)
 if __name__ == "__main__":
-    envs = make_envs(num_envs=16,env_name="CartPole-v0") 
+    envs = make_envs(num_envs=16,env="CartPole-v0") 

codes/A2C/main.py

@@ -5,16 +5,20 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-11 20:58:21
 @LastEditor: John
-LastEditTime: 2021-03-20 16:58:04
+LastEditTime: 2021-04-05 11:14:39
 @Discription: 
 @Environment: python 3.7.9
 '''
 import sys,os
-sys.path.append(os.getcwd()) # add current terminal path
+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 datetime
 from A2C.agent import A2C
+from common.utils import save_results,make_dir,del_empty_dir
 
 
 

codes/A2C/test.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-20 17:43:17
 LastEditor: John
-LastEditTime: 2021-03-20 19:36:24
+LastEditTime: 2021-04-05 11:19:20
 Discription: 
 Environment: 
 '''
@@ -40,7 +40,7 @@ class A2CConfig:
         self.eval_eps = 200
         self.eval_steps = 200
         self.target_update = 4
-        self.hidden_dim=256
+        self.hidden_dim = 256
         self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     
     
@@ -76,7 +76,7 @@ class A2C:
 def train(cfg,env,agent):
     n_states = env.observation_space.shape[0]
     n_actions = env.action_space.n
-    actor_critic = ActorCritic(n_states, n_actions, hidden_dim)
+    actor_critic = ActorCritic(n_states, n_actions, cfg.hidden_dim)
     ac_optimizer = optim.Adam(actor_critic.parameters(), lr=learning_rate)
 
     all_lengths = []
@@ -112,7 +112,7 @@ def train(cfg,env,agent):
                 all_lengths.append(steps)
                 average_lengths.append(np.mean(all_lengths[-10:]))
                 if episode % 10 == 0:                    
-                    sys.stdout.write("episode: {}, reward: {}, total length: {}, average length: {} \n".format(episode, np.sum(rewards), steps, average_lengths[-1]))
+                    sys.stdout.write("episode: {}, reward: {}, total length: {}, average length: {} \n".format(episode, np.sum(rewards), steps+1, average_lengths[-1]))
                 break
         
         # compute Q values
@@ -154,7 +154,7 @@ def train(cfg,env,agent):
     plt.show()
 
 if __name__ == "__main__":
-    cfg = A2CConfig
+    cfg = A2CConfig()
     env = gym.make("CartPole-v0")
     n_states = env.observation_space.shape[0]
     n_actions = env.action_space.n

codes/DDPG/main.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-11 20:58:21
 @LastEditor: John
-LastEditTime: 2021-03-31 01:04:48
+LastEditTime: 2021-04-08 21:50:13
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -35,6 +35,7 @@ if not os.path.exists(RESULT_PATH): os.mkdir(RESULT_PATH)
 
 class DDPGConfig:
     def __init__(self):
+        self.env = 'Pendulum-v0'
         self.algo = 'DDPG'
         self.gamma = 0.99
         self.critic_lr = 1e-3  
@@ -81,6 +82,7 @@ def train(cfg,env,agent):
 
 if __name__ == "__main__":
     cfg = DDPGConfig()
+    env = 
     env = NormalizedActions(gym.make("Pendulum-v0"))
     env.seed(1) # 设置env随机种子
     state_dim = env.observation_space.shape[0]

codes/DQN/main.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:48:57
 @LastEditor: John
-LastEditTime: 2021-04-04 00:26:47
+LastEditTime: 2021-04-13 19:03:39
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -21,15 +21,13 @@ from DQN.agent import DQN
 from common.plot import plot_rewards
 from common.utils import save_results,make_dir,del_empty_dir
 
-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
-RESULT_PATH = curr_path+"/results/"+SEQUENCE+'/' # path to save rewards
-make_dir(curr_path+"/saved_model/",curr_path+"/results/")
-del_empty_dir(curr_path+"/saved_model/",curr_path+"/results/")
+curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # obtain current time
 
 class DQNConfig:
     def __init__(self):
         self.algo = "DQN"  # name of algo
+        self.env = 'CartPole-v0'
+        self.result_path = curr_path+"/results/" +self.env+'/'+curr_time+'/'  # path to save results
         self.gamma = 0.95
         self.epsilon_start = 1 # e-greedy策略的初始epsilon
         self.epsilon_end = 0.01
@@ -37,7 +35,7 @@ class DQNConfig:
         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_eps = 10 # 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
@@ -72,14 +70,13 @@ def train(cfg,env,agent):
 
 if __name__ == "__main__":
     cfg = DQNConfig()
-    env = gym.make('CartPole-v0')
+    env = gym.make(cfg.env)
     env.seed(1)
     state_dim = env.observation_space.shape[0]
     action_dim = env.action_space.n
     agent = DQN(state_dim,action_dim,cfg)
     rewards,ma_rewards = train(cfg,env,agent)
-    make_dir(SAVED_MODEL_PATH,RESULT_PATH)
-    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)
-    del_empty_dir(SAVED_MODEL_PATH,RESULT_PATH)
+    make_dir(cfg.result_path)
+    agent.save(path=cfg.result_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)

codes/DQN/task1.py

@@ -0,0 +1,88 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+@Author: John
+@Email: johnjim0816@gmail.com
+@Date: 2020-06-12 00:48:57
+@LastEditor: John
+LastEditTime: 2021-04-13 18:49:44
+@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 DQN.agent import DQN
+from common.plot import plot_rewards
+from common.utils import save_results,make_dir,del_empty_dir
+
+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
+RESULT_PATH = curr_path+"/results/"+SEQUENCE+'/' # path to save rewards
+make_dir(curr_path+"/saved_model/",curr_path+"/results/")
+del_empty_dir(curr_path+"/saved_model/",curr_path+"/results/")
+
+class DQNConfig:
+    def __init__(self):
+        self.env = 'LunarLander-v2'
+        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 = 500
+        self.lr = 0.0001 # learning rate
+        self.memory_capacity = 1000000 # Replay Memory容量
+        self.batch_size = 64
+        self.train_eps = 300 # 训练的episode数目
+        self.train_steps = 1000
+        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 # 神经网络隐藏层维度
+ 
+def train(cfg,env,agent):
+    print('Start to train !')
+    rewards = []
+    ma_rewards = [] # moveing average reward
+    for i_episode in range(cfg.train_eps):
+        state = env.reset() 
+        ep_reward = 0
+        for i_step in range(cfg.train_steps):
+            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_episode % cfg.target_update == 0:
+            agent.target_net.load_state_dict(agent.policy_net.state_dict())
+        print('Episode:{}/{}, Reward:{}'.format(i_episode+1,cfg.train_eps,ep_reward))
+        rewards.append(ep_reward)
+        # 计算滑动窗口的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 = DQNConfig()
+    env = gym.make(cfg.env)
+    env.seed(1)
+    state_dim = env.observation_space.shape[0]
+    action_dim = env.action_space.n
+    agent = DQN(state_dim,action_dim,cfg)
+    rewards,ma_rewards = train(cfg,env,agent)
+    make_dir(SAVED_MODEL_PATH,RESULT_PATH)
+    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)
+    del_empty_dir(SAVED_MODEL_PATH,RESULT_PATH)

codes/DQN_cnn/main.py

@@ -5,13 +5,11 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-11 10:01:09
 @LastEditor: John
-LastEditTime: 2021-03-29 20:23:48
+LastEditTime: 2021-04-05 11:06:23
 @Discription: 
 @Environment: python 3.7.7
 '''
 import sys,os
-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

codes/PPO/README.md

@@ -0,0 +1,141 @@
+## 原理简介
+PPO是一种off-policy算法，具有较好的性能，其前身是TRPO算法，也是policy gradient算法的一种，它是现在 OpenAI 默认的强化学习算法，具体原理可参考[PPO算法讲解](https://datawhalechina.github.io/easy-rl/#/chapter5/chapter5)。PPO算法主要有两个变种，一个是结合KL penalty的，一个是用了clip方法，本文实现的是后者即```PPO-clip```。
+## 伪代码
+要实现必先了解伪代码，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)
+这是谷歌找到的一张比较适合的图，本人比较懒就没有修改，上面的```k```就是第```k```个episode，第六步是用随机梯度下降的方法优化，这里的损失函数(即```argmax```后面的部分)可能有点难理解，可参考[PPO paper](https://arxiv.org/abs/1707.06347)，如下：
+![在这里插入图片描述](assets/20210323154236878.png)
+第七步就是一个平方损失函数，即实际回报与期望回报的差平方。
+## 代码实战
+[点击查看完整代码](https://github.com/JohnJim0816/rl-tutorials/tree/master/PPO)
+### PPOmemory
+首先第三步需要搜集一条轨迹信息，我们可以定义一个```PPOmemory```来存储相关信息：
+```python
+class PPOMemory:
+    def __init__(self, batch_size):
+        self.states = []
+        self.probs = []
+        self.vals = []
+        self.actions = []
+        self.rewards = []
+        self.dones = []
+        self.batch_size = batch_size
+    def sample(self):
+        batch_step = np.arange(0, len(self.states), self.batch_size)
+        indices = np.arange(len(self.states), dtype=np.int64)
+        np.random.shuffle(indices)
+        batches = [indices[i:i+self.batch_size] for i in batch_step]
+        return np.array(self.states),\
+                np.array(self.actions),\
+                np.array(self.probs),\
+                np.array(self.vals),\
+                np.array(self.rewards),\
+                np.array(self.dones),\
+                batches
+
+    def push(self, state, action, probs, vals, reward, done):
+        self.states.append(state)
+        self.actions.append(action)
+        self.probs.append(probs)
+        self.vals.append(vals)
+        self.rewards.append(reward)
+        self.dones.append(done)
+
+    def clear(self):
+        self.states = []
+        self.probs = []
+        self.actions = []
+        self.rewards = []
+        self.dones = []
+        self.vals = []
+```
+这里的push函数就是将得到的相关量放入memory中，sample就是随机采样出来，方便第六步的随机梯度下降。
+### PPO model
+model就是actor和critic两个网络了：
+```python
+import torch.nn as nn
+from torch.distributions.categorical import Categorical
+class Actor(nn.Module):
+    def __init__(self,state_dim, action_dim,
+            hidden_dim=256):
+        super(Actor, self).__init__()
+
+        self.actor = nn.Sequential(
+                nn.Linear(state_dim, hidden_dim),
+                nn.ReLU(),
+                nn.Linear(hidden_dim, hidden_dim),
+                nn.ReLU(),
+                nn.Linear(hidden_dim, action_dim),
+                nn.Softmax(dim=-1)
+        )
+    def forward(self, state):
+        dist = self.actor(state)
+        dist = Categorical(dist)
+        return dist
+
+class Critic(nn.Module):
+    def __init__(self, state_dim,hidden_dim=256):
+        super(Critic, self).__init__()
+        self.critic = nn.Sequential(
+                nn.Linear(state_dim, hidden_dim),
+                nn.ReLU(),
+                nn.Linear(hidden_dim, hidden_dim),
+                nn.ReLU(),
+                nn.Linear(hidden_dim, 1)
+        )
+    def forward(self, state):
+        value = self.critic(state)
+        return value
+```
+这里Actor就是得到一个概率分布(Categorica，也可以是别的分布，可以搜索torch distributionsl)，critc根据当前状态得到一个值，这里的输入维度可以是```state_dim+action_dim```，即将action信息也纳入critic网络中，这样会更好一些，感兴趣的小伙伴可以试试。
+
+### PPO update
+定义一个update函数主要实现伪代码中的第六步和第七步：
+```python
+def update(self):
+    for _ in range(self.n_epochs):
+        state_arr, action_arr, old_prob_arr, vals_arr,\
+        reward_arr, dones_arr, batches = \
+                self.memory.sample()
+        values = vals_arr
+        ### compute advantage ###
+        advantage = np.zeros(len(reward_arr), dtype=np.float32)
+        for t in range(len(reward_arr)-1):
+            discount = 1
+            a_t = 0
+            for k in range(t, len(reward_arr)-1):
+                a_t += discount*(reward_arr[k] + self.gamma*values[k+1]*\
+                        (1-int(dones_arr[k])) - values[k])
+                discount *= self.gamma*self.gae_lambda
+            advantage[t] = a_t
+        advantage = torch.tensor(advantage).to(self.device)
+        ### SGD ###
+        values = torch.tensor(values).to(self.device)
+        for batch in batches:
+            states = torch.tensor(state_arr[batch], dtype=torch.float).to(self.device)
+            old_probs = torch.tensor(old_prob_arr[batch]).to(self.device)
+            actions = torch.tensor(action_arr[batch]).to(self.device)
+            dist = self.actor(states)
+            critic_value = self.critic(states)
+            critic_value = torch.squeeze(critic_value)
+            new_probs = dist.log_prob(actions)
+            prob_ratio = new_probs.exp() / old_probs.exp()
+            weighted_probs = advantage[batch] * prob_ratio
+            weighted_clipped_probs = torch.clamp(prob_ratio, 1-self.policy_clip,
+                    1+self.policy_clip)*advantage[batch]
+            actor_loss = -torch.min(weighted_probs, weighted_clipped_probs).mean()
+            returns = advantage[batch] + values[batch]
+            critic_loss = (returns-critic_value)**2
+            critic_loss = critic_loss.mean()
+            total_loss = actor_loss + 0.5*critic_loss
+            self.actor_optimizer.zero_grad()
+            self.critic_optimizer.zero_grad()
+            total_loss.backward()
+            self.actor_optimizer.step()
+            self.critic_optimizer.step()
+    self.memory.clear()
+```
+该部分首先从memory中提取搜集到的轨迹信息，然后计算gae，即advantage，接着使用随机梯度下降更新网络，最后清除memory以便搜集下一条轨迹信息。
+
+最后实现效果如下：
+![在这里插入图片描述](assets/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L0pvaG5KaW0w,size_16,color_FFFFFF,t_70-20210405110725113.png)

codes/PPO/agent.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-23 15:17:42
 LastEditor: John
-LastEditTime: 2021-03-23 15:52:34
+LastEditTime: 2021-04-11 01:24:24
 Discription: 
 Environment: 
 '''
@@ -17,16 +17,18 @@ from PPO.model import Actor,Critic
 from PPO.memory import PPOMemory
 class PPO:
     def __init__(self, state_dim, action_dim,cfg):
+        self.env = cfg.env
         self.gamma = cfg.gamma
         self.policy_clip = cfg.policy_clip
         self.n_epochs = cfg.n_epochs
         self.gae_lambda = cfg.gae_lambda
         self.device = cfg.device
-        self.actor = Actor(state_dim, action_dim).to(self.device)
-        self.critic = Critic(state_dim).to(self.device)
-        self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=cfg.lr)
-        self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=cfg.lr)
+        self.actor = Actor(state_dim, action_dim,cfg.hidden_dim).to(self.device)
+        self.critic = Critic(state_dim,cfg.hidden_dim).to(self.device)
+        self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=cfg.actor_lr)
+        self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=cfg.critic_lr)
         self.memory = PPOMemory(cfg.batch_size)
+        self.loss = 0
 
     def choose_action(self, observation):
         state = torch.tensor([observation], dtype=torch.float).to(self.device)
@@ -74,6 +76,7 @@ class PPO:
                 critic_loss = (returns-critic_value)**2
                 critic_loss = critic_loss.mean()
                 total_loss = actor_loss + 0.5*critic_loss
+                self.loss  = total_loss
                 self.actor_optimizer.zero_grad()
                 self.critic_optimizer.zero_grad()
                 total_loss.backward()
@@ -81,13 +84,13 @@ class PPO:
                 self.critic_optimizer.step()
         self.memory.clear()  
     def save(self,path):
-        actor_checkpoint = os.path.join(path, 'actor_torch_ppo.pt')
-        critic_checkpoint= os.path.join(path, 'critic_torch_ppo.pt')
+        actor_checkpoint = os.path.join(path, self.env+'_actor.pt')
+        critic_checkpoint= os.path.join(path, self.env+'_critic.pt')
         torch.save(self.actor.state_dict(), actor_checkpoint)
         torch.save(self.critic.state_dict(), critic_checkpoint)
     def load(self,path):
-        actor_checkpoint = os.path.join(path, 'actor_torch_ppo.pt')
-        critic_checkpoint= os.path.join(path, 'critic_torch_ppo.pt')
+        actor_checkpoint = os.path.join(path, self.env+'_actor.pt')
+        critic_checkpoint= os.path.join(path, self.env+'_critic.pt')
         self.actor.load_state_dict(torch.load(actor_checkpoint)) 
         self.critic.load_state_dict(torch.load(critic_checkpoint))  
 

codes/PPO/main.py

@@ -5,12 +5,14 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-22 16:18:10
 LastEditor: John
-LastEditTime: 2021-03-23 15:52:52
+LastEditTime: 2021-04-11 01:24:41
 Discription: 
 Environment: 
 '''
 import sys,os
-sys.path.append(os.getcwd()) # add current terminal path to sys.path
+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 numpy as np
 import torch
@@ -33,15 +35,18 @@ if not os.path.exists(RESULT_PATH): # 检测是否存在文件夹
 
 class PPOConfig:
     def __init__(self) -> None:
+        self.env = 'CartPole-v0'
         self.algo = 'PPO'
         self.batch_size = 5
         self.gamma=0.99
         self.n_epochs = 4
-        self.lr = 0.0003
+        self.actor_lr = 0.0003
+        self.critic_lr = 0.0003
         self.gae_lambda=0.95
         self.policy_clip=0.2
+        self.hidden_dim = 256
         self.update_fre = 20 # frequency of agent update
-        self.train_eps = 250 # max training episodes
+        self.train_eps = 300 # max training episodes
         self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # check gpu
         
 def train(cfg,env,agent):
@@ -70,7 +75,8 @@ def train(cfg,env,agent):
         else:
             ma_rewards.append(ep_reward)
         avg_reward = np.mean(rewards[-100:])
-        if avg_reward > best_reward:
+        if avg_rewardself.actor_lr = 0.002
+        self.critic_lr = 0.005 > best_reward:
             best_reward = avg_reward
             agent.save(path=SAVED_MODEL_PATH)
         print('Episode:{}/{}, Reward:{:.1f}, avg reward:{:.1f}, Done:{}'.format(i_episode+1,cfg.train_eps,ep_reward,avg_reward,done))
@@ -78,7 +84,7 @@ def train(cfg,env,agent):
 
 if __name__ == '__main__':
     cfg = PPOConfig()
-    env = gym.make('CartPole-v0')
+    env = gym.make(cfg.env)
     env.seed(1)
     state_dim=env.observation_space.shape[0]
     action_dim=env.action_space.n

codes/PPO/model.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-23 15:29:24
 LastEditor: John
-LastEditTime: 2021-03-23 15:29:52
+LastEditTime: 2021-04-08 22:36:43
 Discription: 
 Environment: 
 '''
@@ -13,7 +13,7 @@ import torch.nn as nn
 from torch.distributions.categorical import Categorical
 class Actor(nn.Module):
     def __init__(self,state_dim, action_dim,
-            hidden_dim=256):
+            hidden_dim):
         super(Actor, self).__init__()
 
         self.actor = nn.Sequential(
@@ -30,7 +30,7 @@ class Actor(nn.Module):
         return dist
 
 class Critic(nn.Module):
-    def __init__(self, state_dim,hidden_dim=256):
+    def __init__(self, state_dim,hidden_dim):
         super(Critic, self).__init__()
         self.critic = nn.Sequential(
                 nn.Linear(state_dim, hidden_dim),

codes/PPO/task1.py

@@ -0,0 +1,97 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2021-03-22 16:18:10
+LastEditor: John
+LastEditTime: 2021-04-11 01:25:43
+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 gym
+import numpy as np
+import torch
+import datetime
+from PPO.agent import PPO
+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)
+
+class PPOConfig:
+    def __init__(self) -> None:
+        self.env = 'LunarLander-v2'
+        self.algo = 'PPO'
+        self.batch_size = 128
+        self.gamma=0.95
+        self.n_epochs = 4
+        self.actor_lr = 0.002
+        self.critic_lr = 0.005
+        self.gae_lambda=0.95
+        self.policy_clip=0.2
+        self.hidden_dim = 256
+        self.update_fre = 20 # frequency of agent update
+        self.train_eps = 300 # max training episodes
+        self.train_steps = 1000
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # check gpu
+        
+def train(cfg,env,agent):
+    best_reward = env.reward_range[0]
+    rewards= []
+    ma_rewards = [] # moving average rewards
+    avg_reward = 0
+    running_steps = 0
+    for i_episode in range(cfg.train_eps):
+        state = env.reset()
+        done = False
+        ep_reward = 0
+        # for i_step in range(cfg.train_steps):
+        while not done:
+            action, prob, val = agent.choose_action(state)
+            state_, reward, done, _ = env.step(action)
+            running_steps += 1
+            ep_reward += reward
+            agent.memory.push(state, action, prob, val, reward, done)
+            if running_steps % cfg.update_fre == 0:
+                agent.update()
+            state = state_
+            # if done:
+            #     break
+        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)
+        avg_reward = np.mean(rewards[-100:])
+        if avg_reward > best_reward:
+            best_reward = avg_reward
+            agent.save(path=SAVED_MODEL_PATH)
+        print('Episode:{}/{}, Reward:{:.1f}, avg reward:{:.1f}, Loss:{}'.format(i_episode+1,cfg.train_eps,ep_reward,avg_reward,agent.loss))
+    return rewards,ma_rewards
+
+if __name__ == '__main__':
+    cfg = PPOConfig()
+    env = gym.make(cfg.env)
+    env.seed(1)
+    state_dim=env.observation_space.shape[0]
+    action_dim=env.action_space.n
+    agent = PPO(state_dim,action_dim,cfg)
+    rewards,ma_rewards = train(cfg,env,agent)
+    save_results(rewards,ma_rewards,tag='train',path=RESULT_PATH)
+    plot_rewards(rewards,ma_rewards,tag="train",algo = cfg.algo,path=RESULT_PATH)

codes/README.md

@@ -22,26 +22,26 @@ python 3.7、pytorch 1.6.0-1.7.1、gym 0.17.0-0.18.0
 
 ## 使用说明
 
-运行```main.py```或者```main.ipynb```
+运行```main.py```或者```main.ipynb```，或者包含```task```名的文件(比如```task1.py```)
 ## 算法进度
 
-|                 算法名称                 |                        相关论文材料                         | 环境                                  |                备注                |
-| :--------------------------------------: | :---------------------------------------------------------: | ------------------------------------- | :--------------------------------: |
-| [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 |
-|         [DoubleDQN](./DoubleDQN)         |                                                             | [CartPole-v0](./envs/gym_info.md)     |                                    |
-|   [Hierarchical DQN](HierarchicalDQN)    |       [H-DQN Paper](https://arxiv.org/abs/1604.06057)       | [CartPole-v0](./envs/gym_info.md)     |                                    |
-|    [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)              |       [DDPG Paper](https://arxiv.org/abs/1509.02971)        | [Pendulum-v0](./envs/gym_info.md)     |                                    |
-|                   TD3                    |        [TD3 Paper](https://arxiv.org/abs/1802.09477)        |                                       |                                    |
-|                   GAIL                   |                                                             |                                       |                                    |
+|                 算法名称                 |                         相关论文材料                         | 环境                                  |                备注                |
+| :--------------------------------------: | :----------------------------------------------------------: | ------------------------------------- | :--------------------------------: |
+| [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),[Nature DQN Paper](https://www.nature.com/articles/nature14236) | [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)     |                                    |
+|   [Hierarchical DQN](HierarchicalDQN)    |       [H-DQN Paper](https://arxiv.org/abs/1604.06057)        | [CartPole-v0](./envs/gym_info.md)     |                                    |
+|    [PolicyGradient](./PolicyGradient)    |                                                              | [CartPole-v0](./envs/gym_info.md)     |                                    |
+|                   A2C                    |        [A3C Paper](https://arxiv.org/abs/1602.01783)         | [CartPole-v0](./envs/gym_info.md)     |                                    |
+|                   A3C                    |        [A3C Paper](https://arxiv.org/abs/1602.01783)         |                                       |                                    |
+|                   SAC                    |        [SAC Paper](https://arxiv.org/abs/1801.01290)         |                                       |                                    |
+|               [PPO](./PPO)               |        [PPO paper](https://arxiv.org/abs/1707.06347)         | [CartPole-v0](./envs/gym_info.md)     |                                    |
+|              [DDPG](./DDPG)              |        [DDPG Paper](https://arxiv.org/abs/1509.02971)        | [Pendulum-v0](./envs/gym_info.md)     |                                    |
+|               [TD3](./TD3)               |        [TD3 Paper](https://arxiv.org/abs/1802.09477)         | HalfCheetah-v2                        |                                    |
+|                   GAIL                   |        [GAIL Paper](https://arxiv.org/abs/1606.03476)        |                                       |                                    |
 
 
 
@@ -53,4 +53,3 @@ python 3.7、pytorch 1.6.0-1.7.1、gym 0.17.0-0.18.0
 
 [RL-Adventure](https://github.com/higgsfield/RL-Adventure)
 
-https://www.cnblogs.com/lucifer1997/p/13458563.html

codes/README_en.md

@@ -22,27 +22,27 @@ 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
 
-run ```main.py``` or ```main.ipynb```
+run ```main.py``` or ```main.ipynb```, or run files with ```task```(like ```task1.py```)
 
 ## Schedule
 
-|                   Name                   |                      Related materials                      | Used Envs                             |  Notes   |
-| :--------------------------------------: | :---------------------------------------------------------: | ------------------------------------- | :------: |
-| [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)     |          |
-|         [DoubleDQN](./DoubleDQN)         |                                                             | [CartPole-v0](./envs/gym_info.md)     | not well |
-|   [Hierarchical DQN](HierarchicalDQN)    |    [Hierarchical DQN](https://arxiv.org/abs/1604.06057)     | [CartPole-v0](./envs/gym_info.md)     |          |
-|    [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)              |       [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) |                                       |          |
-|                   GAIL                   |                                                             |                                       |          |
+|                   Name                   |                      Related materials                      | Used Envs                             | Notes |
+| :--------------------------------------: | :---------------------------------------------------------: | ------------------------------------- | :---: |
+| [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)     |       |
+|         [DoubleDQN](./DoubleDQN)         |                                                             | [CartPole-v0](./envs/gym_info.md)     |       |
+|   [Hierarchical DQN](HierarchicalDQN)    |    [Hierarchical DQN](https://arxiv.org/abs/1604.06057)     | [CartPole-v0](./envs/gym_info.md)     |       |
+|    [PolicyGradient](./PolicyGradient)    |                                                             | [CartPole-v0](./envs/gym_info.md)     |       |
+|                   A2C                    |        [A3C Paper](https://arxiv.org/abs/1602.01783)        | [CartPole-v0](./envs/gym_info.md)     |       |
+|                   A3C                    |        [A3C Paper](https://arxiv.org/abs/1602.01783)        |                                       |       |
+|                   SAC                    |        [SAC Paper](https://arxiv.org/abs/1801.01290)        |                                       |       |
+|               [PPO](./PPO)               |        [PPO paper](https://arxiv.org/abs/1707.06347)        | [CartPole-v0](./envs/gym_info.md)     |       |
+|              [DDPG](./DDPG)              |       [DDPG Paper](https://arxiv.org/abs/1509.02971)        | [Pendulum-v0](./envs/gym_info.md)     |       |
+|               [TD3](./TD3)               |        [TD3 Paper](https://arxiv.org/abs/1802.09477)        | HalfCheetah-v2                        |       |
+|                   GAIL                   |                                                             |                                       |       |
 
 
 ## Refs
@@ -52,4 +52,4 @@ run ```main.py``` or ```main.ipynb```
 
 [RL-Adventure](https://github.com/higgsfield/RL-Adventure)
 
-https://www.cnblogs.com/lucifer1997/p/13458563.html
+

codes/TD3/agent.py

@@ -0,0 +1,170 @@
+import copy
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from TD3.memory import ReplayBuffer
+
+
+
+# Implementation of Twin Delayed Deep Deterministic Policy Gradients (TD3)
+# Paper: https://arxiv.org/abs/1802.09477
+
+
+class Actor(nn.Module):
+	def __init__(self, state_dim, action_dim, max_action):
+		super(Actor, self).__init__()
+
+		self.l1 = nn.Linear(state_dim, 256)
+		self.l2 = nn.Linear(256, 256)
+		self.l3 = nn.Linear(256, action_dim)
+		
+		self.max_action = max_action
+		
+
+	def forward(self, state):
+		a = F.relu(self.l1(state))
+		a = F.relu(self.l2(a))
+		return self.max_action * torch.tanh(self.l3(a))
+
+
+class Critic(nn.Module):
+	def __init__(self, state_dim, action_dim):
+		super(Critic, self).__init__()
+
+		# Q1 architecture
+		self.l1 = nn.Linear(state_dim + action_dim, 256)
+		self.l2 = nn.Linear(256, 256)
+		self.l3 = nn.Linear(256, 1)
+
+		# Q2 architecture
+		self.l4 = nn.Linear(state_dim + action_dim, 256)
+		self.l5 = nn.Linear(256, 256)
+		self.l6 = nn.Linear(256, 1)
+
+
+	def forward(self, state, action):
+		sa = torch.cat([state, action], 1)
+
+		q1 = F.relu(self.l1(sa))
+		q1 = F.relu(self.l2(q1))
+		q1 = self.l3(q1)
+
+		q2 = F.relu(self.l4(sa))
+		q2 = F.relu(self.l5(q2))
+		q2 = self.l6(q2)
+		return q1, q2
+
+
+	def Q1(self, state, action):
+		sa = torch.cat([state, action], 1)
+
+		q1 = F.relu(self.l1(sa))
+		q1 = F.relu(self.l2(q1))
+		q1 = self.l3(q1)
+		return q1
+
+
+class TD3(object):
+	def __init__(
+		self,
+		state_dim,
+		action_dim,
+		max_action,
+		cfg,
+	):
+		self.max_action = max_action
+		self.gamma = cfg.gamma
+		self.lr = cfg.lr
+		self.policy_noise = cfg.policy_noise
+		self.noise_clip = cfg.noise_clip
+		self.policy_freq = cfg.policy_freq
+		self.batch_size =  cfg.batch_size 
+		self.device = cfg.device
+		self.total_it = 0
+
+		self.actor = Actor(state_dim, action_dim, max_action).to(self.device)
+		self.actor_target = copy.deepcopy(self.actor)
+		self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=3e-4)
+
+		self.critic = Critic(state_dim, action_dim).to(self.device)
+		self.critic_target = copy.deepcopy(self.critic)
+		self.critic_optimizer = torch.optim.Adam(self.critic.parameters(), lr=3e-4)
+		self.memory = ReplayBuffer(state_dim, action_dim)
+
+		
+
+
+	def choose_action(self, state):
+		state = torch.FloatTensor(state.reshape(1, -1)).to(self.device)
+		return self.actor(state).cpu().data.numpy().flatten()
+
+
+	def update(self):
+		self.total_it += 1
+
+		# Sample replay buffer 
+		state, action, next_state, reward, not_done = self.memory.sample(self.batch_size)
+
+		with torch.no_grad():
+			# Select action according to policy and add clipped noise
+			noise = (
+				torch.randn_like(action) * self.policy_noise
+			).clamp(-self.noise_clip, self.noise_clip)
+			
+			next_action = (
+				self.actor_target(next_state) + noise
+			).clamp(-self.max_action, self.max_action)
+
+			# Compute the target Q value
+			target_Q1, target_Q2 = self.critic_target(next_state, next_action)
+			target_Q = torch.min(target_Q1, target_Q2)
+			target_Q = reward + not_done * self.gamma * target_Q
+
+		# Get current Q estimates
+		current_Q1, current_Q2 = self.critic(state, action)
+
+		# Compute critic loss
+		critic_loss = F.mse_loss(current_Q1, target_Q) + F.mse_loss(current_Q2, target_Q)
+
+		# Optimize the critic
+		self.critic_optimizer.zero_grad()
+		critic_loss.backward()
+		self.critic_optimizer.step()
+
+		# Delayed policy updates
+		if self.total_it % self.policy_freq == 0:
+
+			# Compute actor losse
+			actor_loss = -self.critic.Q1(state, self.actor(state)).mean()
+			
+			# Optimize the actor 
+			self.actor_optimizer.zero_grad()
+			actor_loss.backward()
+			self.actor_optimizer.step()
+
+			# Update the frozen target models
+			for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()):
+				target_param.data.copy_(self.lr * param.data + (1 - self.lr) * target_param.data)
+
+			for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()):
+				target_param.data.copy_(self.lr * param.data + (1 - self.lr) * target_param.data)
+
+
+	def save(self, path):
+		torch.save(self.critic.state_dict(), path + "td3_critic")
+		torch.save(self.critic_optimizer.state_dict(), path + "td3_critic_optimizer")
+		
+		torch.save(self.actor.state_dict(), path + "td3_actor")
+		torch.save(self.actor_optimizer.state_dict(), path + "td3_actor_optimizer")
+
+
+	def load(self, path):
+		self.critic.load_state_dict(torch.load(path + "td3_critic"))
+		self.critic_optimizer.load_state_dict(torch.load(path + "td3_critic_optimizer"))
+		self.critic_target = copy.deepcopy(self.critic)
+
+		self.actor.load_state_dict(torch.load(path + "td3_actor"))
+		self.actor_optimizer.load_state_dict(torch.load(path + "td3_actor_optimizer"))
+		self.actor_target = copy.deepcopy(self.actor)
+		

codes/TD3/main.py

@@ -1,14 +1,169 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-@Author: John
-@Email: johnjim0816@gmail.com
-@Date: 2020-06-11 23:38:13
-@LastEditor: John
-@LastEditTime: 2020-06-11 23:38:31
-@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 torch
+import gym
+import numpy as np
+import datetime
+
+
+from TD3.agent import TD3
+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 TD3Config:
+	def __init__(self) -> None:
+		self.algo = 'TD3'
+		self.env = 'HalfCheetah-v2'
+		self.seed = 0
+		self.result_path = curr_path+"/results/" +self.env+'/'+curr_time+'/'  # path to save results
+		self.start_timestep = 25e3 # Time steps initial random policy is used
+		self.eval_freq = 5e3 # How often (time steps) we evaluate
+		# self.train_eps = 800
+		self.max_timestep = 1600000 # Max time steps to run environment
+		self.expl_noise = 0.1 # Std of Gaussian exploration noise
+		self.batch_size = 256 # Batch size for both actor and critic
+		self.gamma = 0.99 # gamma factor
+		self.lr = 0.0005 # Target network update rate 
+		self.policy_noise = 0.2 # Noise added to target policy during critic update
+		self.noise_clip = 0.5  # Range to clip target policy noise
+		self.policy_freq = 2 # Frequency of delayed policy updates
+		self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+		
+# Runs policy for X episodes and returns average reward
+# A fixed seed is used for the eval environment
+def eval(env,agent, seed, eval_episodes=10):
+	eval_env = gym.make(env)
+	eval_env.seed(seed + 100)
+	avg_reward = 0.
+	for _ in range(eval_episodes):
+		state, done = eval_env.reset(), False
+		while not done:
+			# eval_env.render()
+			action = agent.choose_action(np.array(state))
+			state, reward, done, _ = eval_env.step(action)
+			avg_reward += reward
+	avg_reward /= eval_episodes
+	print("---------------------------------------")
+	print(f"Evaluation over {eval_episodes} episodes: {avg_reward:.3f}")
+	print("---------------------------------------")
+	return avg_reward
+
+def train(cfg,env,agent):
+	# Evaluate untrained policy
+	evaluations = [eval(cfg.env,agent, cfg.seed)]
+	state, done = env.reset(), False
+	ep_reward = 0
+	ep_timesteps = 0
+	episode_num = 0
+	rewards = []
+	ma_rewards = [] # moveing average reward
+	for t in range(int(cfg.max_timestep)):
+		ep_timesteps += 1
+		# Select action randomly or according to policy
+		if t < cfg.start_timestep:
+			action = env.action_space.sample()
+		else:
+			action = (
+				agent.choose_action(np.array(state))
+				+ np.random.normal(0, max_action * cfg.expl_noise, size=action_dim)
+			).clip(-max_action, max_action)
+		# Perform action
+		next_state, reward, done, _ = env.step(action) 
+		done_bool = float(done) if ep_timesteps < env._max_episode_steps else 0
+		# Store data in replay buffer
+		agent.memory.push(state, action, next_state, reward, done_bool)
+		state = next_state
+		ep_reward += reward
+		# Train agent after collecting sufficient data
+		if t >= cfg.start_timestep:
+			agent.update()
+		if done: 
+			# +1 to account for 0 indexing. +0 on ep_timesteps since it will increment +1 even if done=True
+			print(f"Episode:{episode_num+1}, Episode T:{ep_timesteps}, Reward:{ep_reward:.3f}")
+			# Reset environment
+			state, done = env.reset(), False
+			rewards.append(ep_reward)
+			# 计算滑动窗口的reward
+			if ma_rewards:
+				ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
+			else:
+				ma_rewards.append(ep_reward) 
+			ep_reward = 0
+			ep_timesteps = 0
+			episode_num += 1 
+		# Evaluate episode
+		if (t + 1) % cfg.eval_freq == 0:
+			evaluations.append(eval(cfg.env,agent, cfg.seed))
+	return rewards, ma_rewards
+# def train(cfg,env,agent):
+# 	evaluations = [eval(cfg.env,agent,cfg.seed)]
+# 	ep_reward = 0
+# 	tot_timestep = 0
+# 	rewards = []
+# 	ma_rewards = [] # moveing average reward
+# 	for i_ep in range(int(cfg.train_eps)):
+# 		state, done = env.reset(), False
+# 		ep_reward = 0
+# 		ep_timestep = 0
+# 		while not done:
+# 			ep_timestep += 1
+# 			tot_timestep +=1
+# 			# Select action randomly or according to policy
+# 			if tot_timestep < cfg.start_timestep:
+# 				action = env.action_space.sample()
+# 			else:
+# 				action = (
+# 					agent.choose_action(np.array(state))
+# 					+ np.random.normal(0, max_action * cfg.expl_noise, size=action_dim)
+# 				).clip(-max_action, max_action)
+# 			# action = (
+# 			# 		agent.choose_action(np.array(state))
+# 			# 		+ np.random.normal(0, max_action * cfg.expl_noise, size=action_dim)
+# 			# 	).clip(-max_action, max_action)
+# 			# Perform action
+# 			next_state, reward, done, _ = env.step(action) 
+# 			done_bool = float(done) if ep_timestep < env._max_episode_steps else 0
+
+# 			# Store data in replay buffer
+# 			agent.memory.push(state, action, next_state, reward, done_bool)
+# 			state = next_state
+# 			ep_reward += reward
+# 			# Train agent after collecting sufficient data
+# 			if tot_timestep >= cfg.start_timestep:
+# 				agent.update()
+# 		print(f"Episode:{i_ep}/{cfg.train_eps}, Episode Timestep:{ep_timestep}, Reward:{ep_reward:.3f}")
+# 		rewards.append(ep_reward)
+# 		# 计算滑动窗口的reward
+# 		if ma_rewards:
+# 			ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
+# 		else:
+# 			ma_rewards.append(ep_reward) 
+# 		# Evaluate episode
+# 		if (i_ep+1) % cfg.eval_freq == 0:
+# 			evaluations.append(eval(cfg.env,agent, cfg.seed))
+# 	return rewards,ma_rewards
+
+
 if __name__ == "__main__":
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+	cfg  = TD3Config()
+	env = gym.make(cfg.env)
+	env.seed(cfg.seed) # Set seeds
+	torch.manual_seed(cfg.seed)
+	np.random.seed(cfg.seed)
+	state_dim = env.observation_space.shape[0]
+	action_dim = env.action_space.shape[0] 
+	max_action = float(env.action_space.high[0])
+	agent = TD3(state_dim,action_dim,max_action,cfg)
+	rewards,ma_rewards = train(cfg,env,agent)
+	make_dir(cfg.result_path)
+	agent.save(path=cfg.result_path)
+	save_results(rewards,ma_rewards,tag='train',path=cfg.result_path)
+	plot_rewards(rewards,ma_rewards,tag="train",env=cfg.env,algo = cfg.algo,path=cfg.result_path)
+	
+		

codes/TD3/memory.py

@@ -1,34 +1,44 @@
 #!/usr/bin/env python
 # coding=utf-8
 '''
-@Author: John
-@Email: johnjim0816@gmail.com
-@Date: 2020-06-10 15:27:16
-@LastEditor: John
-@LastEditTime: 2020-06-11 21:04:50
-@Discription: 
-@Environment: python 3.7.7
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2021-04-13 11:00:13
+LastEditor: John
+LastEditTime: 2021-04-15 01:25:14
+Discription: 
+Environment: 
 '''
-import random
 import numpy as np
+import torch
 
-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 = map(np.stack, zip(*batch))
-        return state, action, reward, next_state, done
-    
-    def __len__(self):
-        return len(self.buffer)
+
+class ReplayBuffer(object):
+	def __init__(self, state_dim, action_dim, max_size=int(1e6)):
+		self.max_size = max_size
+		self.ptr = 0
+		self.size = 0
+		self.state = np.zeros((max_size, state_dim))
+		self.action = np.zeros((max_size, action_dim))
+		self.next_state = np.zeros((max_size, state_dim))
+		self.reward = np.zeros((max_size, 1))
+		self.not_done = np.zeros((max_size, 1))
+		self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+	def push(self, state, action, next_state, reward, done):
+		self.state[self.ptr] = state
+		self.action[self.ptr] = action
+		self.next_state[self.ptr] = next_state
+		self.reward[self.ptr] = reward
+		self.not_done[self.ptr] = 1. - done
+		self.ptr = (self.ptr + 1) % self.max_size
+		self.size = min(self.size + 1, self.max_size)
+
+	def sample(self, batch_size):
+		ind = np.random.randint(0, self.size, size=batch_size)
+		return (
+			torch.FloatTensor(self.state[ind]).to(self.device),
+			torch.FloatTensor(self.action[ind]).to(self.device),
+			torch.FloatTensor(self.next_state[ind]).to(self.device),
+			torch.FloatTensor(self.reward[ind]).to(self.device),
+			torch.FloatTensor(self.not_done[ind]).to(self.device)
+		)

codes/common/plot.py

@@ -5,15 +5,15 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-10-07 20:57:11
 LastEditor: John
-LastEditTime: 2021-03-31 18:47:28
+LastEditTime: 2021-04-08 21:45:09
 Discription: 
 Environment: 
 '''
 import matplotlib.pyplot as plt
 import seaborn as sns
-def plot_rewards(rewards,ma_rewards,tag="train",algo = "DQN",save=True,path='./'):
+def plot_rewards(rewards,ma_rewards,tag="train",env='CartPole-v0',algo = "DQN",save=True,path='./'):
     sns.set()
-    plt.title("average learning curve of {}".format(algo))
+    plt.title("average learning curve of {} for {}".format(algo,env))
     plt.xlabel('epsiodes')
     plt.plot(rewards,label='rewards')
     plt.plot(ma_rewards,label='moving average rewards')

codes/common/utils.py

@@ -5,12 +5,14 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-12 16:02:24
 LastEditor: John
-LastEditTime: 2021-04-03 21:42:13
+LastEditTime: 2021-04-13 18:34:20
 Discription: 
 Environment: 
 '''
 import os
 import numpy as np
+from pathlib import Path
+
 
 
 def save_results(rewards,ma_rewards,tag='train',path='./results'):
@@ -22,8 +24,7 @@ def save_results(rewards,ma_rewards,tag='train',path='./results'):
 
 def make_dir(*paths):
     for path in paths:
-        if not os.path.exists(path): # check if exists
-            os.mkdir(path)
+        Path(path).mkdir(parents=True, exist_ok=True)
 def del_empty_dir(*paths):
     '''del_empty_dir delete empty folders unders "paths"
     '''

codes/envs/snake/main.py

@@ -10,7 +10,7 @@ import time
 def get_args():
     parser = argparse.ArgumentParser(description='CS440 MP4 Snake')
 
-    parser.add_argument('--human', default = True, action="store_true",
+    parser.add_argument('--human', default = False, action="store_true",
                         help='making the game human playable - default False')
 
     parser.add_argument('--model_name', dest="model_name", type=str, default="checkpoint3.npy",

codes/test.py

@@ -1,19 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-Author: John
-Email: johnjim0816@gmail.com
-Date: 2021-03-25 23:25:15
-LastEditor: John
-LastEditTime: 2021-03-26 16:46:52
-Discription: 
-Environment: 
-'''
-from collections import defaultdict
-import numpy as np
-action_dim = 2
-Q_table  = defaultdict(lambda: np.zeros(action_dim))
-Q_table[str(0)] = 1
-print(Q_table[str(0)])
-Q_table[str(21)] = 3
-print(Q_table[str(21)])