update

codes/DDPG/task0.py

@@ -34,7 +34,7 @@ class DDPGConfig:
         self.env_name = env_name # 环境名称
         self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测GPU
         self.train_eps = 300 # 训练的回合数
-        self.eval_eps = 50 # 测试的回合数
+        self.test_eps = 50 # 测试的回合数
         self.gamma = 0.99 # 折扣因子
         self.critic_lr = 1e-3 # 评论家网络的学习率
         self.actor_lr = 1e-4 # 演员网络的学习率

codes/DDPG/train.py

@@ -42,7 +42,7 @@ def test(cfg, env, agent):
     print(f'环境：{cfg.env_name}, 算法：{cfg.algo}, 设备：{cfg.device}')
     rewards = [] # 记录所有回合的奖励
     ma_rewards = []  # 记录所有回合的滑动平均奖励
-    for i_ep in range(cfg.eval_eps):
+    for i_ep in range(cfg.test_eps):
         state = env.reset() 
         done = False
         ep_reward = 0
@@ -59,6 +59,6 @@ def test(cfg, env, agent):
             ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
         else:
             ma_rewards.append(ep_reward)
-        print(f"回合：{i_ep+1}/{cfg.eval_eps}，奖励：{ep_reward:.1f}")
+        print(f"回合：{i_ep+1}/{cfg.test_eps}，奖励：{ep_reward:.1f}")
     print('完成测试！')
     return rewards, ma_rewards

codes/DQN/task0.py

@@ -23,7 +23,7 @@ class DQNConfig:
         self.device = torch.device(
             "cuda" if torch.cuda.is_available() else "cpu")  # 检测GPU
         self.train_eps = 200  # 训练的回合数
-        self.eval_eps = 30  # 测试的回合数
+        self.test_eps = 30  # 测试的回合数
         # 超参数
         self.gamma = 0.95  # 强化学习中的折扣因子
         self.epsilon_start = 0.90  # e-greedy策略中初始epsilon

codes/DQN/task1.py

@@ -26,7 +26,7 @@ class DQNConfig:
         self.device = torch.device(
             "cuda" if torch.cuda.is_available() else "cpu")  # 检测GPU
         self.train_eps = 200  # 训练的回合数
-        self.eval_eps = 30  # 测试的回合数
+        self.test_eps = 30  # 测试的回合数
         # 超参数
         self.gamma = 0.95  # 强化学习中的折扣因子
         self.epsilon_start = 0.90  # e-greedy策略中初始epsilon

codes/DQN/train.ipynb

@@ -180,7 +180,7 @@
     "        self.algo = \"DQN\"  # 算法名称\n",
     "        self.env = 'CartPole-v0' # 环境名称\n",
     "        self.train_eps = 200 # 训练的回合数\n",
-    "        self.eval_eps = 20 # 测试的回合数\n",
+    "        self.test_eps = 20 # 测试的回合数\n",
     "        self.gamma = 0.95 # 强化学习中的折扣因子\n",
     "        self.epsilon_start = 0.90 # e-greedy策略中初始epsilon\n",
     "        self.epsilon_end = 0.01 # e-greedy策略中的终止epsilon\n",
@@ -365,7 +365,7 @@
     "    cfg.epsilon_end = 0.0 # e-greedy策略中的终止epsilon\n",
     "    rewards = [] # 记录所有回合的奖励\n",
     "    ma_rewards = []  # 记录所有回合的滑动平均奖励\n",
-    "    for i_ep in range(cfg.eval_eps):\n",
+    "    for i_ep in range(cfg.test_eps):\n",
     "        ep_reward = 0 # 记录一回合内的奖励\n",
     "        state = env.reset() # 重置环境，返回初始状态\n",
     "        while True:\n",
@@ -381,7 +381,7 @@
     "        else:\n",
     "            ma_rewards.append(ep_reward)\n",
     "        if (i_ep+1)%3 == 0: \n",
-    "            print(f\"回合：{i_ep+1}/{cfg.eval_eps}, 奖励：{ep_reward:.1f}\")\n",
+    "            print(f\"回合：{i_ep+1}/{cfg.test_eps}, 奖励：{ep_reward:.1f}\")\n",
     "    print('完成测试！')\n",
     "    return rewards,ma_rewards\n",
     "\n",

codes/DQN/train.py

@@ -5,7 +5,7 @@
 @Email: johnjim0816@gmail.com
 @Date: 2020-06-12 00:48:57
 @LastEditor: John
-LastEditTime: 2021-09-15 15:34:13
+LastEditTime: 2021-12-22 11:08:04
 @Discription: 
 @Environment: python 3.7.7
 '''
@@ -30,13 +30,13 @@ def train(cfg, env, agent):
                 break
         if (i_ep+1) % cfg.target_update == 0: # 智能体目标网络更新
             agent.target_net.load_state_dict(agent.policy_net.state_dict())
-        if (i_ep+1)%10 == 0: 
-            print('回合：{}/{}, 奖励：{}'.format(i_ep+1, cfg.train_eps, ep_reward))
         rewards.append(ep_reward)
         if ma_rewards:
             ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
         else:
             ma_rewards.append(ep_reward)
+        if (i_ep+1)%10 == 0: 
+            print('回合：{}/{}, 奖励：{}'.format(i_ep+1, cfg.train_eps, ep_reward))
     print('完成训练！')
     return rewards, ma_rewards
 
@@ -48,7 +48,7 @@ def test(cfg,env,agent):
     cfg.epsilon_end = 0.0 # e-greedy策略中的终止epsilon
     rewards = [] # 记录所有回合的奖励
     ma_rewards = []  # 记录所有回合的滑动平均奖励
-    for i_ep in range(cfg.eval_eps):
+    for i_ep in range(cfg.test_eps):
         ep_reward = 0 # 记录一回合内的奖励
         state = env.reset() # 重置环境，返回初始状态
         while True:
@@ -63,7 +63,7 @@ def test(cfg,env,agent):
             ma_rewards.append(ma_rewards[-1]*0.9+ep_reward*0.1)
         else:
             ma_rewards.append(ep_reward)
-        print(f"回合：{i_ep+1}/{cfg.eval_eps}，奖励：{ep_reward:.1f}")
+        print(f"回合：{i_ep+1}/{cfg.test_eps}，奖励：{ep_reward:.1f}")
     print('完成测试！')
     return rewards,ma_rewards
 
@@ -89,7 +89,7 @@ if __name__ == "__main__":
             self.env_name = 'CartPole-v0' # 环境名称
             self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")  # 检测GPU
             self.train_eps = 200 # 训练的回合数
-            self.eval_eps = 30 # 测试的回合数
+            self.test_eps = 30 # 测试的回合数
             # 超参数
             self.gamma = 0.95 # 强化学习中的折扣因子
             self.epsilon_start = 0.90 # e-greedy策略中初始epsilon

codes/PPO/task0.py

@@ -20,7 +20,7 @@ class PPOConfig:
         self.continuous = False # 环境是否为连续动作
         self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")  # 检测GPU
         self.train_eps = 200 # 训练的回合数
-        self.eval_eps = 20 # 测试的回合数
+        self.test_eps = 20 # 测试的回合数
         self.batch_size = 5
         self.gamma=0.99
         self.n_epochs = 4

codes/PPO/task1.py

@@ -20,7 +20,7 @@ class PPOConfig:
         self.continuous = True # 环境是否为连续动作
         self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")  # 检测GPU
         self.train_eps = 200 # 训练的回合数
-        self.eval_eps = 20 # 测试的回合数
+        self.test_eps = 20 # 测试的回合数
         self.batch_size = 5
         self.gamma=0.99
         self.n_epochs = 4

codes/PPO/train.ipynb

@@ -68,7 +68,7 @@
     "        self.result_path = curr_path+\"/results/\" +self.env+'/'+curr_time+'/results/'  # path to save results\n",
     "        self.model_path = curr_path+\"/results/\" +self.env+'/'+curr_time+'/models/'  # path to save models\n",
     "        self.train_eps = 200 # max training episodes\n",
-    "        self.eval_eps = 50\n",
+    "        self.test_eps = 50\n",
     "        self.batch_size = 5\n",
     "        self.gamma=0.99\n",
     "        self.n_epochs = 4\n",
@@ -144,7 +144,7 @@
     "    print(f'Env:{cfg.env}, Algorithm:{cfg.algo}, Device:{cfg.device}')\n",
     "    rewards= []\n",
     "    ma_rewards = [] # moving average rewards\n",
-    "    for i_ep in range(cfg.eval_eps):\n",
+    "    for i_ep in range(cfg.test_eps):\n",
     "        state = env.reset()\n",
     "        done = False\n",
     "        ep_reward = 0\n",

codes/PPO/train.py

@@ -32,7 +32,7 @@ def eval(cfg,env,agent):
     print(f'环境：{cfg.env_name}, 算法：{cfg.algo}, 设备：{cfg.device}')
     rewards = [] # 记录所有回合的奖励
     ma_rewards = []  # 记录所有回合的滑动平均奖励
-    for i_ep in range(cfg.eval_eps):
+    for i_ep in range(cfg.test_eps):
         state = env.reset()
         done = False
         ep_reward = 0
@@ -47,7 +47,7 @@ def eval(cfg,env,agent):
                 0.9*ma_rewards[-1]+0.1*ep_reward)
         else:
             ma_rewards.append(ep_reward)
-        print('回合：{}/{}, 奖励：{}'.format(i_ep+1, cfg.eval_eps, ep_reward))
+        print('回合：{}/{}, 奖励：{}'.format(i_ep+1, cfg.test_eps, ep_reward))
     print('完成训练！')
     return rewards,ma_rewards
 
@@ -74,7 +74,7 @@ if __name__ == '__main__':
             self.continuous = False # 环境是否为连续动作
             self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")  # 检测GPU
             self.train_eps = 200 # 训练的回合数
-            self.eval_eps = 20 # 测试的回合数
+            self.test_eps = 20 # 测试的回合数
             self.batch_size = 5
             self.gamma=0.99
             self.n_epochs = 4

codes/PolicyGradient/model.py

@@ -5,21 +5,22 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2021-03-23 16:35:58
 LastEditor: John
-LastEditTime: 2021-03-23 16:36:20
+LastEditTime: 2021-12-21 23:21:26
 Discription: 
 Environment: 
 '''
 import torch.nn as nn
 import torch.nn.functional as F
 class MLP(nn.Module):
+    
     ''' 多层感知机
         输入：state维度
         输出：概率
     '''
-    def __init__(self,state_dim,hidden_dim = 36):
+    def __init__(self,input_dim,hidden_dim = 36):
         super(MLP, self).__init__()
-        # 24和36为hidden layer的层数，可根据state_dim, action_dim的情况来改变
-        self.fc1 = nn.Linear(state_dim, hidden_dim)
+        # 24和36为hidden layer的层数，可根据input_dim, action_dim的情况来改变
+        self.fc1 = nn.Linear(input_dim, hidden_dim)
         self.fc2 = nn.Linear(hidden_dim,hidden_dim)
         self.fc3 = nn.Linear(hidden_dim, 1)  # Prob of Left
 

codes/PolicyGradient/task0_train.py

@@ -34,7 +34,7 @@ class PGConfig:
         self.model_path = curr_path+"/outputs/" + self.env + \
             '/'+curr_time+'/models/'  # 保存模型的路径
         self.train_eps = 300 # 训练的回合数
-        self.eval_eps = 30 # 测试的回合数
+        self.test_eps = 30 # 测试的回合数
         self.batch_size = 8
         self.lr = 0.01 # 学习率
         self.gamma = 0.99
@@ -94,7 +94,7 @@ def eval(cfg,env,agent):
     print(f'Env:{cfg.env}, Algorithm:{cfg.algo}, Device:{cfg.device}')
     rewards = []
     ma_rewards = []
-    for i_ep in range(cfg.eval_eps):
+    for i_ep in range(cfg.test_eps):
         state = env.reset()
         ep_reward = 0
         for _ in count():

codes/QLearning/agent.py

@@ -5,7 +5,7 @@ Author: John
 Email: johnjim0816@gmail.com
 Date: 2020-09-11 23:03:00
 LastEditor: John
-LastEditTime: 2021-09-19 23:05:45
+LastEditTime: 2021-12-22 10:54:57
 Discription: use defaultdict to define Q table
 Environment: 
 '''
@@ -15,17 +15,17 @@ import torch
 from collections import defaultdict
 
 class QLearning(object):
-    def __init__(self,state_dim,
-                 action_dim,cfg):
-        self.action_dim = action_dim  # dimension of acgtion
-        self.lr = cfg.lr  # learning rate
+    def __init__(self,n_states,
+                 n_actions,cfg):
+        self.n_actions = n_actions 
+        self.lr = cfg.lr  # 学习率
         self.gamma = cfg.gamma  
         self.epsilon = 0 
         self.sample_count = 0  
         self.epsilon_start = cfg.epsilon_start
         self.epsilon_end = cfg.epsilon_end
         self.epsilon_decay = cfg.epsilon_decay
-        self.Q_table  = defaultdict(lambda: np.zeros(action_dim)) # A nested dictionary that maps state -> (action -> action-value)
+        self.Q_table  = defaultdict(lambda: np.zeros(n_actions)) # 用嵌套字典存放状态->动作->状态-动作值（Q值）的映射，即Q表
     def choose_action(self, state):
         self.sample_count += 1
         self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * \
@@ -34,7 +34,7 @@ class QLearning(object):
         if np.random.uniform(0, 1) > self.epsilon:
             action = np.argmax(self.Q_table[str(state)]) # 选择Q(s,a)最大对应的动作
         else:
-            action = np.random.choice(self.action_dim) # 随机选择动作
+            action = np.random.choice(self.n_actions) # 随机选择动作
         return action
     def predict(self,state):
         action = np.argmax(self.Q_table[str(state)])

codes/QLearning/task0.py

@@ -0,0 +1,93 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2020-09-11 23:03:00
+LastEditor: John
+LastEditTime: 2021-12-22 11:13:23
+Discription: 
+Environment: 
+'''
+import sys
+import os
+curr_path = os.path.dirname(os.path.abspath(__file__))  # 当前文件所在绝对路径
+parent_path = os.path.dirname(curr_path)  # 父路径
+sys.path.append(parent_path)  # 添加路径到系统路径
+
+import gym
+import torch
+import datetime
+
+from envs.gridworld_env import CliffWalkingWapper
+from QLearning.agent import QLearning
+from QLearning.train import train,test
+from common.utils import plot_rewards,plot_rewards_cn
+from common.utils import save_results,make_dir
+
+curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # 获取当前时间
+algo_name = 'Q-learning'  # 算法名称
+env_name = 'CliffWalking-v0'  # 环境名称
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测GPU
+class QlearningConfig:
+    '''训练相关参数'''
+    def __init__(self):
+        self.algo_name = algo_name # 算法名称
+        self.env_name = env_name # 环境名称
+        self.device = device # 检测GPU
+        self.train_eps = 400 # 训练的回合数
+        self.test_eps = 30 # 测试的回合数
+        self.gamma = 0.9 # reward的衰减率
+        self.epsilon_start = 0.95 # e-greedy策略中初始epsilon
+        self.epsilon_end = 0.01 # e-greedy策略中的终止epsilon
+        self.epsilon_decay = 300 # e-greedy策略中epsilon的衰减率
+        self.lr = 0.1 # 学习率      
+class PlotConfig:
+    ''' 绘图相关参数设置
+    '''
+
+    def __init__(self) -> None:
+        self.algo_name = algo_name  # 算法名称
+        self.env_name = env_name  # 环境名称
+        self.device = device # 检测GPU
+        self.result_path = curr_path + "/outputs/" + self.env_name + \
+            '/' + curr_time + '/results/'  # 保存结果的路径
+        self.model_path = curr_path + "/outputs/" + self.env_name + \
+            '/' + curr_time + '/models/'  # 保存模型的路径
+        self.save = True  # 是否保存图片
+        
+def env_agent_config(cfg,seed=1):
+    '''创建环境和智能体
+    Args:
+        cfg ([type]): [description]
+        seed (int, optional): 随机种子. Defaults to 1.
+    Returns:
+        env [type]: 环境
+        agent : 智能体
+    '''    
+    env = gym.make(cfg.env_name)  
+    env = CliffWalkingWapper(env)
+    env.seed(seed) # 设置随机种子
+    n_states = env.observation_space.n # 状态维度
+    n_actions = env.action_space.n # 动作维度
+    agent = QLearning(n_states,n_actions,cfg)
+    return env,agent
+
+cfg = QlearningConfig()
+plot_cfg = PlotConfig()
+# 训练
+env, agent = env_agent_config(cfg, seed=1)
+rewards, ma_rewards = train(cfg, env, agent)
+make_dir(plot_cfg.result_path, plot_cfg.model_path)  # 创建保存结果和模型路径的文件夹
+agent.save(path=plot_cfg.model_path)  # 保存模型
+save_results(rewards, ma_rewards, tag='train',
+            path=plot_cfg.result_path)  # 保存结果
+plot_rewards(rewards, ma_rewards, plot_cfg, tag="train")  # 画出结果
+# 测试
+env, agent = env_agent_config(cfg, seed=10)
+agent.load(path=plot_cfg.model_path)  # 导入模型
+rewards, ma_rewards = test(cfg, env, agent)
+save_results(rewards, ma_rewards, tag='test', path=plot_cfg.result_path)  # 保存结果
+plot_rewards(rewards, ma_rewards, plot_cfg, tag="test")  # 画出结果
+        
+    

codes/QLearning/task0_train.py

@@ -1,126 +0,0 @@
-#!/usr/bin/env python
-# coding=utf-8
-'''
-Author: John
-Email: johnjim0816@gmail.com
-Date: 2020-09-11 23:03:00
-LastEditor: John
-LastEditTime: 2021-09-23 12:22:58
-Discription: 
-Environment: 
-'''
-import sys,os
-curr_path = os.path.dirname(os.path.abspath(__file__)) # 当前路径
-parent_path=os.path.dirname(curr_path) # 父路径，这里就是我们的项目路径
-sys.path.append(parent_path) # 由于需要引用项目路径下的其他模块比如envs，所以需要添加路径到sys.path
-
-import gym
-import torch
-import datetime
-
-from envs.gridworld_env import CliffWalkingWapper
-from QLearning.agent import QLearning
-from common.plot import plot_rewards,plot_rewards_cn
-from common.utils import save_results,make_dir
-
-curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") # 获取当前时间
-class QlearningConfig:
-    '''训练相关参数'''
-    def __init__(self):
-        self.algo = 'Q-learning' # 算法名称
-        self.env = 'CliffWalking-v0' # 环境名称
-        self.result_path = curr_path+"/outputs/" +self.env+'/'+curr_time+'/results/'  # 保存结果的路径
-        self.model_path = curr_path+"/outputs/" +self.env+'/'+curr_time+'/models/'  # 保存模型的路径
-        self.train_eps = 400 # 训练的回合数
-        self.eval_eps = 30 # 测试的回合数
-        self.gamma = 0.9 # reward的衰减率
-        self.epsilon_start = 0.95 # e-greedy策略中初始epsilon
-        self.epsilon_end = 0.01 # e-greedy策略中的终止epsilon
-        self.epsilon_decay = 300 # e-greedy策略中epsilon的衰减率
-        self.lr = 0.1 # 学习率
-        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 检测GPU
-
-        
-def env_agent_config(cfg,seed=1):
-    env = gym.make(cfg.env)  
-    env = CliffWalkingWapper(env)
-    env.seed(seed) # 设置随机种子
-    state_dim = env.observation_space.n # 状态维度
-    action_dim = env.action_space.n # 动作维度
-    agent = QLearning(state_dim,action_dim,cfg)
-    return env,agent
-
-def train(cfg,env,agent):
-    print('开始训练！')
-    print(f'环境:{cfg.env}, 算法:{cfg.algo}, 设备:{cfg.device}')
-    rewards = []  # 记录奖励
-    ma_rewards = [] # 记录滑动平均奖励
-    for i_ep in range(cfg.train_eps):
-        ep_reward = 0  # 记录每个回合的奖励
-        state = env.reset()  # 重置环境,即开始新的回合
-        while True:
-            action = agent.choose_action(state)  # 根据算法选择一个动作
-            next_state, reward, done, _ = env.step(action)  # 与环境进行一次动作交互
-            print(reward)
-            agent.update(state, action, reward, next_state, done)  # Q学习算法更新
-            state = next_state  # 更新状态
-            ep_reward += reward
-            if done:
-                break
-        rewards.append(ep_reward)
-        if ma_rewards:
-            ma_rewards.append(ma_rewards[-1]*0.9+ep_reward*0.1)
-        else:
-            ma_rewards.append(ep_reward)
-        print("回合数：{}/{}，奖励{:.1f}".format(i_ep+1, cfg.train_eps,ep_reward))
-    print('完成训练！')
-    return rewards,ma_rewards
-    
-def eval(cfg,env,agent):
-    print('开始测试！')
-    print(f'环境：{cfg.env}, 算法：{cfg.algo}, 设备：{cfg.device}')
-    for item in agent.Q_table.items():
-        print(item)
-    rewards = []  # 记录所有回合的奖励
-    ma_rewards = [] # 滑动平均的奖励
-    for i_ep in range(cfg.eval_eps):
-        ep_reward = 0  # 记录每个episode的reward
-        state = env.reset()  # 重置环境, 重新开一局（即开始新的一个回合）
-        while True:
-            action = agent.predict(state)  # 根据算法选择一个动作
-            next_state, reward, done, _ = env.step(action)  # 与环境进行一个交互
-            state = next_state  # 更新状态
-            ep_reward += reward
-            if done:
-                break
-        rewards.append(ep_reward)
-        if ma_rewards:
-            ma_rewards.append(ma_rewards[-1]*0.9+ep_reward*0.1)
-        else:
-            ma_rewards.append(ep_reward)
-        print(f"回合数：{i_ep+1}/{cfg.eval_eps}, 奖励：{ep_reward:.1f}")
-    print('完成测试！')
-    return rewards,ma_rewards
-    
-if __name__ == "__main__":
-    cfg = QlearningConfig()
-
-    # 训练
-    env,agent = env_agent_config(cfg,seed=0)
-    rewards,ma_rewards = train(cfg,env,agent)
-    make_dir(cfg.result_path,cfg.model_path) # 创建文件夹
-    agent.save(path=cfg.model_path) # 保存模型
-    for item in agent.Q_table.items():
-        print(item)
-    save_results(rewards,ma_rewards,tag='train',path=cfg.result_path) # 保存结果
-    plot_rewards_cn(rewards,ma_rewards,tag="train",env=cfg.env,algo = cfg.algo,path=cfg.result_path)
-
-    # # 测试
-    env,agent = env_agent_config(cfg,seed=10)
-    agent.load(path=cfg.model_path) # 加载模型
-    rewards,ma_rewards = eval(cfg,env,agent)
-    
-    save_results(rewards,ma_rewards,tag='eval',path=cfg.result_path)
-    plot_rewards_cn(rewards,ma_rewards,tag="eval",env=cfg.env,algo = cfg.algo,path=cfg.result_path)
-    
-    

codes/QLearning/train.py

@@ -0,0 +1,51 @@
+def train(cfg,env,agent):
+    print('开始训练！')
+    print(f'环境:{cfg.env_name}, 算法:{cfg.algo_name}, 设备:{cfg.device}')
+    rewards = []  # 记录奖励
+    ma_rewards = [] # 记录滑动平均奖励
+    for i_ep in range(cfg.train_eps):
+        ep_reward = 0  # 记录每个回合的奖励
+        state = env.reset()  # 重置环境,即开始新的回合
+        while True:
+            action = agent.choose_action(state)  # 根据算法选择一个动作
+            next_state, reward, done, _ = env.step(action)  # 与环境进行一次动作交互
+            agent.update(state, action, reward, next_state, done)  # Q学习算法更新
+            state = next_state  # 更新状态
+            ep_reward += reward
+            if done:
+                break
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(ma_rewards[-1]*0.9+ep_reward*0.1)
+        else:
+            ma_rewards.append(ep_reward)
+        if ()
+        print("回合数：{}/{}，奖励{:.1f}".format(i_ep+1, cfg.train_eps,ep_reward))
+    print('完成训练！')
+    return rewards,ma_rewards
+    
+def test(cfg,env,agent):
+    print('开始测试！')
+    print(f'环境：{cfg.env_name}, 算法：{cfg.algo_name}, 设备：{cfg.device}')
+    for item in agent.Q_table.items():
+        print(item)
+    rewards = []  # 记录所有回合的奖励
+    ma_rewards = [] # 滑动平均的奖励
+    for i_ep in range(cfg.test_eps):
+        ep_reward = 0  # 记录每个episode的reward
+        state = env.reset()  # 重置环境, 重新开一局（即开始新的一个回合）
+        while True:
+            action = agent.predict(state)  # 根据算法选择一个动作
+            next_state, reward, done, _ = env.step(action)  # 与环境进行一个交互
+            state = next_state  # 更新状态
+            ep_reward += reward
+            if done:
+                break
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(ma_rewards[-1]*0.9+ep_reward*0.1)
+        else:
+            ma_rewards.append(ep_reward)
+        print(f"回合数：{i_ep+1}/{cfg.test_eps}, 奖励：{ep_reward:.1f}")
+    print('完成测试！')
+    return rewards,ma_rewards

codes/README.md

@@ -13,6 +13,7 @@
 
 其中```model.py```,```memory.py```,```plot.py``` 由于不同算法都会用到，所以放入```common```文件夹中。
 
+**注意：新版本中将```model```,```memory```相关内容全部放到了```agent.py```里面，```plot```放到了```common.utils```中。**
 ## 运行环境
 
 python 3.7、pytorch 1.6.0-1.8.1、gym 0.17.0-0.19.0

codes/SAC/task0_train.ipynb

@@ -45,7 +45,7 @@
     "        self.model_path = curr_path+\"/outputs/\" +self.env+'/'+curr_time+'/models/'  # path to save models\n",
     "        self.train_eps = 300\n",
     "        self.train_steps = 500\n",
-    "        self.eval_eps = 50\n",
+    "        self.test_eps = 50\n",
     "        self.eval_steps = 500\n",
     "        self.gamma = 0.99\n",
     "        self.mean_lambda=1e-3\n",
@@ -121,7 +121,7 @@
     "    print(f'Env: {cfg.env}, Algorithm: {cfg.algo}, Device: {cfg.device}')\n",
     "    rewards  = []\n",
     "    ma_rewards = [] # moveing average reward\n",
-    "    for i_ep in range(cfg.eval_eps):\n",
+    "    for i_ep in range(cfg.test_eps):\n",
     "        state = env.reset()\n",
     "        ep_reward = 0\n",
     "        for i_step in range(cfg.eval_steps):\n",

codes/SAC/task0_train.py

@@ -33,7 +33,7 @@ class SACConfig:
         self.model_path = curr_path+"/outputs/" +self.env_name+'/'+curr_time+'/models/'  # path to save models
         self.train_eps = 300
         self.train_steps = 500
-        self.eval_eps = 50
+        self.test_eps = 50
         self.eval_steps = 500
         self.gamma = 0.99
         self.mean_lambda=1e-3
@@ -96,7 +96,7 @@ def eval(cfg,env,agent):
     print(f'Env: {cfg.env_name}, Algorithm: {cfg.algo}, Device: {cfg.device}')
     rewards  = []
     ma_rewards = [] # moveing average reward
-    for i_ep in range(cfg.eval_eps):
+    for i_ep in range(cfg.test_eps):
         state = env.reset()
         ep_reward = 0
         for i_step in range(cfg.eval_steps):

codes/Sarsa/task0_train.py

@@ -31,7 +31,7 @@ class SarsaConfig:
         self.result_path = curr_path+"/outputs/" +self.env+'/'+curr_time+'/results/'  # path to save results
         self.model_path = curr_path+"/outputs/" +self.env+'/'+curr_time+'/models/'  # path to save models
         self.train_eps = 200
-        self.eval_eps = 50
+        self.test_eps = 50
         self.epsilon = 0.15 # epsilon: The probability to select a random action . 
         self.gamma = 0.9 # gamma: Gamma discount factor.
         self.lr = 0.2 # learning rate: step size parameter
@@ -74,7 +74,7 @@ def train(cfg,env,agent):
 def eval(cfg,env,agent):
     rewards = []
     ma_rewards = []
-    for i_episode in range(cfg.eval_eps):
+    for i_episode in range(cfg.test_eps):
         # Print out which episode we're on, useful for debugging.
         # Generate an episode.
         # An episode is an array of (state, action, reward) tuples
@@ -94,7 +94,7 @@ def eval(cfg,env,agent):
             ma_rewards.append(ep_reward)
         rewards.append(ep_reward)
         if (i_episode+1)%10==0:
-            print("Episode:{}/{}: Reward:{}".format(i_episode+1, cfg.eval_eps,ep_reward))
+            print("Episode:{}/{}: Reward:{}".format(i_episode+1, cfg.test_eps,ep_reward))
     print('Complete evaling！')
     return rewards,ma_rewards
         

codes/TD3/README.md

@@ -0,0 +1 @@
+这是对[Implementation of Twin Delayed Deep Deterministic Policy Gradients (TD3)](https://arxiv.org/abs/1802.09477)的复现

codes/TD3/agent.py

@@ -1,3 +1,13 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: JiangJi
+Email: johnjim0816@gmail.com
+Date: 2021-12-22 10:40:05
+LastEditor: JiangJi
+LastEditTime: 2021-12-22 10:43:55
+Discription: 
+'''
 import copy
 import numpy as np
 import torch
@@ -5,40 +15,41 @@ 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):
+	
+	def __init__(self, input_dim, output_dim, max_action):
+		'''[summary]
+
+		Args:
+			input_dim (int): 输入维度，这里等于n_states
+			output_dim (int): 输出维度，这里等于n_actions
+			max_action (int): action的最大值
+		'''		
 		super(Actor, self).__init__()
 
-		self.l1 = nn.Linear(state_dim, 256)
+		self.l1 = nn.Linear(input_dim, 256)
 		self.l2 = nn.Linear(256, 256)
-		self.l3 = nn.Linear(256, action_dim)
-		
+		self.l3 = nn.Linear(256, output_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):
+	def __init__(self, input_dim, output_dim):
 		super(Critic, self).__init__()
 
 		# Q1 architecture
-		self.l1 = nn.Linear(state_dim + action_dim, 256)
+		self.l1 = nn.Linear(input_dim + output_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.l4 = nn.Linear(input_dim + output_dim, 256)
 		self.l5 = nn.Linear(256, 256)
 		self.l6 = nn.Linear(256, 1)
 
@@ -68,8 +79,8 @@ class Critic(nn.Module):
 class TD3(object):
 	def __init__(
 		self,
-		state_dim,
-		action_dim,
+		input_dim,
+		output_dim,
 		max_action,
 		cfg,
 	):
@@ -83,14 +94,14 @@ class TD3(object):
 		self.device = cfg.device
 		self.total_it = 0
 
-		self.actor = Actor(state_dim, action_dim, max_action).to(self.device)
+		self.actor = Actor(input_dim, output_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 = Critic(input_dim, output_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)
+		self.memory = ReplayBuffer(input_dim, output_dim)
 
 	def choose_action(self, state):
 		state = torch.FloatTensor(state.reshape(1, -1)).to(self.device)