update projects

projects/codes/A2C/README.md

@@ -0,0 +1,5 @@
+## A2C
+
+
+
+https://towardsdatascience.com/understanding-actor-critic-methods-931b97b6df3f

projects/codes/A2C/a2c.py

@@ -0,0 +1,56 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: JiangJi
+Email: johnjim0816@gmail.com
+Date: 2021-05-03 22:16:08
+LastEditor: JiangJi
+LastEditTime: 2022-07-20 23:54:40
+Discription: 
+Environment: 
+'''
+import torch
+import torch.optim as optim
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.distributions import Categorical
+
+class ActorCritic(nn.Module):
+    ''' A2C网络模型，包含一个Actor和Critic
+    '''
+    def __init__(self, input_dim, output_dim, hidden_dim):
+        super(ActorCritic, self).__init__()
+        self.critic = nn.Sequential(
+            nn.Linear(input_dim, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, 1)
+        )
+        
+        self.actor = nn.Sequential(
+            nn.Linear(input_dim, hidden_dim),
+            nn.ReLU(),
+            nn.Linear(hidden_dim, output_dim),
+            nn.Softmax(dim=1),
+        )
+        
+    def forward(self, x):
+        value = self.critic(x)
+        probs = self.actor(x)
+        dist  = Categorical(probs)
+        return dist, value
+class A2C:
+    ''' A2C算法
+    '''
+    def __init__(self,n_states,n_actions,cfg) -> None:
+        self.gamma = cfg.gamma
+        self.device = torch.device(cfg.device)
+        self.model = ActorCritic(n_states, n_actions, cfg.hidden_size).to(self.device)
+        self.optimizer = optim.Adam(self.model.parameters())
+
+    def compute_returns(self,next_value, rewards, masks):
+        R = next_value
+        returns = []
+        for step in reversed(range(len(rewards))):
+            R = rewards[step] + self.gamma * R * masks[step]
+            returns.insert(0, R)
+        return returns

projects/codes/A2C/outputs/CartPole-v0/20220713-221850/results/params.json

@@ -0,0 +1,14 @@
+{
+    "algo_name": "A2C",
+    "env_name": "CartPole-v0",
+    "n_envs": 8,
+    "max_steps": 20000,
+    "n_steps": 5,
+    "gamma": 0.99,
+    "lr": 0.001,
+    "hidden_dim": 256,
+    "deivce": "cpu",
+    "result_path": "C:\\Users\\24438\\Desktop\\rl-tutorials/outputs/CartPole-v0/20220713-221850/results/",
+    "model_path": "C:\\Users\\24438\\Desktop\\rl-tutorials/outputs/CartPole-v0/20220713-221850/models/",
+    "save_fig": true
+}

projects/codes/A2C/task0.py

@@ -0,0 +1,137 @@
+import sys,os
+curr_path = os.path.dirname(os.path.abspath(__file__))  # current path
+parent_path = os.path.dirname(curr_path)  # parent path
+sys.path.append(parent_path)  # add to system path
+
+import gym
+import numpy as np
+import torch
+import torch.optim as optim
+import datetime
+import argparse
+from common.multiprocessing_env import SubprocVecEnv
+from a2c import ActorCritic
+from common.utils import save_results, make_dir
+from common.utils import plot_rewards, save_args
+
+
+def get_args():
+    """ Hyperparameters
+    """
+    curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")  # Obtain current time
+    parser = argparse.ArgumentParser(description="hyperparameters")      
+    parser.add_argument('--algo_name',default='A2C',type=str,help="name of algorithm")
+    parser.add_argument('--env_name',default='CartPole-v0',type=str,help="name of environment")
+    parser.add_argument('--n_envs',default=8,type=int,help="numbers of environments")
+
+    parser.add_argument('--max_steps',default=20000,type=int,help="episodes of training")
+    parser.add_argument('--n_steps',default=5,type=int,help="episodes of testing")
+    parser.add_argument('--gamma',default=0.99,type=float,help="discounted factor")
+    parser.add_argument('--lr',default=1e-3,type=float,help="learning rate")
+    parser.add_argument('--hidden_dim',default=256,type=int)
+    parser.add_argument('--device',default='cpu',type=str,help="cpu or cuda") 
+    parser.add_argument('--result_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/results/' )
+    parser.add_argument('--model_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/models/' ) # path to save models
+    parser.add_argument('--save_fig',default=True,type=bool,help="if save figure or not")           
+    args = parser.parse_args()                        
+    return args
+
+def make_envs(env_name):
+    def _thunk():
+        env = gym.make(env_name)
+        env.seed(2)
+        return env
+    return _thunk
+def test_env(env,model,vis=False):
+    state = env.reset()
+    if vis: env.render()
+    done = False
+    total_reward = 0
+    while not done:
+        state = torch.FloatTensor(state).unsqueeze(0).to(cfg.device)
+        dist, _ = model(state)
+        next_state, reward, done, _ = env.step(dist.sample().cpu().numpy()[0])
+        state = next_state
+        if vis: env.render()
+        total_reward += reward
+    return total_reward
+
+def compute_returns(next_value, rewards, masks, gamma=0.99):
+    R = next_value
+    returns = []
+    for step in reversed(range(len(rewards))):
+        R = rewards[step] + gamma * R * masks[step]
+        returns.insert(0, R)
+    return returns
+
+
+def train(cfg,envs):
+    print('Start training!')
+    print(f'Env:{cfg.env_name}, Algorithm:{cfg.algo_name}, Device:{cfg.device}')
+    env = gym.make(cfg.env_name) # a single env
+    env.seed(10)
+    n_states  = envs.observation_space.shape[0]
+    n_actions = envs.action_space.n
+    model = ActorCritic(n_states, n_actions, cfg.hidden_dim).to(cfg.device)
+    optimizer = optim.Adam(model.parameters())
+    step_idx    = 0
+    test_rewards = []
+    test_ma_rewards = []
+    state = envs.reset()
+    while step_idx < cfg.max_steps:
+        log_probs = []
+        values    = []
+        rewards   = []
+        masks     = []
+        entropy = 0
+        # rollout trajectory
+        for _ in range(cfg.n_steps):
+            state = torch.FloatTensor(state).to(cfg.device)
+            dist, value = model(state)
+            action = dist.sample()
+            next_state, reward, done, _ = envs.step(action.cpu().numpy())
+            log_prob = dist.log_prob(action)
+            entropy += dist.entropy().mean()
+            log_probs.append(log_prob)
+            values.append(value)
+            rewards.append(torch.FloatTensor(reward).unsqueeze(1).to(cfg.device))
+            masks.append(torch.FloatTensor(1 - done).unsqueeze(1).to(cfg.device))
+            state = next_state
+            step_idx += 1
+            if step_idx % 100 == 0:
+                test_reward = np.mean([test_env(env,model) for _ in range(10)])
+                print(f"step_idx:{step_idx}, test_reward:{test_reward}")
+                test_rewards.append(test_reward)
+                if test_ma_rewards:
+                    test_ma_rewards.append(0.9*test_ma_rewards[-1]+0.1*test_reward)
+                else:
+                    test_ma_rewards.append(test_reward) 
+                # plot(step_idx, test_rewards)   
+        next_state = torch.FloatTensor(next_state).to(cfg.device)
+        _, next_value = model(next_state)
+        returns = compute_returns(next_value, rewards, masks)
+        log_probs = torch.cat(log_probs)
+        returns   = torch.cat(returns).detach()
+        values    = torch.cat(values)
+        advantage = returns - values
+        actor_loss  = -(log_probs * advantage.detach()).mean()
+        critic_loss = advantage.pow(2).mean()
+        loss = actor_loss + 0.5 * critic_loss - 0.001 * entropy
+        optimizer.zero_grad()
+        loss.backward()
+        optimizer.step()
+    print('Finish training！')
+    return {'rewards':test_rewards,'ma_rewards':test_ma_rewards}
+if __name__ == "__main__":
+    cfg = get_args()
+    envs = [make_envs(cfg.env_name) for i in range(cfg.n_envs)]
+    envs = SubprocVecEnv(envs) 
+    # training
+    res_dic = train(cfg,envs)
+    make_dir(cfg.result_path,cfg.model_path)
+    save_args(cfg)
+    save_results(res_dic, tag='train',
+                 path=cfg.result_path)
+    plot_rewards(res_dic['rewards'], res_dic['ma_rewards'], cfg, tag="train") # 画出结果

projects/codes/DDPG/ddpg.py

@@ -0,0 +1,144 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+@Author: John
+@Email: johnjim0816@gmail.com
+@Date: 2020-06-09 20:25:52
+@LastEditor: John
+LastEditTime: 2022-06-09 19:04:44
+@Discription: 
+@Environment: python 3.7.7
+'''
+import random
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import torch.nn.functional as F
+class ReplayBuffer:
+    def __init__(self, capacity):
+        self.capacity = capacity # 经验回放的容量
+        self.buffer = [] # 缓冲区
+        self.position = 0 
+    
+    def push(self, state, action, reward, next_state, done):
+        ''' 缓冲区是一个队列，容量超出时去掉开始存入的转移(transition)
+        '''
+        if len(self.buffer) < self.capacity:
+            self.buffer.append(None)
+        self.buffer[self.position] = (state, action, reward, next_state, done)
+        self.position = (self.position + 1) % self.capacity 
+    
+    def sample(self, batch_size):
+        batch = random.sample(self.buffer, batch_size) # 随机采出小批量转移
+        state, action, reward, next_state, done =  zip(*batch) # 解压成状态，动作等
+        return state, action, reward, next_state, done
+    
+    def __len__(self):
+        ''' 返回当前存储的量
+        '''
+        return len(self.buffer)
+class Actor(nn.Module):
+    def __init__(self, n_states, n_actions, hidden_dim, init_w=3e-3):
+        super(Actor, self).__init__()  
+        self.linear1 = nn.Linear(n_states, hidden_dim)
+        self.linear2 = nn.Linear(hidden_dim, hidden_dim)
+        self.linear3 = nn.Linear(hidden_dim, n_actions)
+        
+        self.linear3.weight.data.uniform_(-init_w, init_w)
+        self.linear3.bias.data.uniform_(-init_w, init_w)
+        
+    def forward(self, x):
+        x = F.relu(self.linear1(x))
+        x = F.relu(self.linear2(x))
+        x = torch.tanh(self.linear3(x))
+        return x
+class Critic(nn.Module):
+    def __init__(self, n_states, n_actions, hidden_dim, init_w=3e-3):
+        super(Critic, self).__init__()
+        
+        self.linear1 = nn.Linear(n_states + n_actions, hidden_dim)
+        self.linear2 = nn.Linear(hidden_dim, hidden_dim)
+        self.linear3 = nn.Linear(hidden_dim, 1)
+        # 随机初始化为较小的值
+        self.linear3.weight.data.uniform_(-init_w, init_w)
+        self.linear3.bias.data.uniform_(-init_w, init_w)
+        
+    def forward(self, state, action):
+        # 按维数1拼接
+        x = torch.cat([state, action], 1)
+        x = F.relu(self.linear1(x))
+        x = F.relu(self.linear2(x))
+        x = self.linear3(x)
+        return x
+class DDPG:
+    def __init__(self, n_states, n_actions, cfg):
+        self.device = torch.device(cfg.device)
+        self.critic = Critic(n_states, n_actions, cfg.hidden_dim).to(self.device)
+        self.actor = Actor(n_states, n_actions, cfg.hidden_dim).to(self.device)
+        self.target_critic = Critic(n_states, n_actions, cfg.hidden_dim).to(self.device)
+        self.target_actor = Actor(n_states, n_actions, cfg.hidden_dim).to(self.device)
+
+        # 复制参数到目标网络
+        for target_param, param in zip(self.target_critic.parameters(), self.critic.parameters()):
+            target_param.data.copy_(param.data)
+        for target_param, param in zip(self.target_actor.parameters(), self.actor.parameters()):
+            target_param.data.copy_(param.data)
+
+        self.critic_optimizer = optim.Adam(
+            self.critic.parameters(),  lr=cfg.critic_lr)
+        self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=cfg.actor_lr)
+        self.memory = ReplayBuffer(cfg.memory_capacity)
+        self.batch_size = cfg.batch_size
+        self.soft_tau = cfg.soft_tau # 软更新参数
+        self.gamma = cfg.gamma
+
+    def choose_action(self, state):
+        state = torch.FloatTensor(state).unsqueeze(0).to(self.device)
+        action = self.actor(state)
+        return action.detach().cpu().numpy()[0, 0]
+
+    def update(self):
+        if len(self.memory) < self.batch_size: # 当 memory 中不满足一个批量时，不更新策略
+            return
+        # 从经验回放中(replay memory)中随机采样一个批量的转移(transition)
+        state, action, reward, next_state, done = self.memory.sample(self.batch_size)
+        # 转变为张量
+        state = torch.FloatTensor(np.array(state)).to(self.device)
+        next_state = torch.FloatTensor(np.array(next_state)).to(self.device)
+        action = torch.FloatTensor(np.array(action)).to(self.device)
+        reward = torch.FloatTensor(reward).unsqueeze(1).to(self.device)
+        done = torch.FloatTensor(np.float32(done)).unsqueeze(1).to(self.device)
+       
+        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
+        expected_value = torch.clamp(expected_value, -np.inf, np.inf)
+
+        value = self.critic(state, action)
+        value_loss = nn.MSELoss()(value, expected_value.detach())
+        
+        self.actor_optimizer.zero_grad()
+        policy_loss.backward()
+        self.actor_optimizer.step()
+        self.critic_optimizer.zero_grad()
+        value_loss.backward()
+        self.critic_optimizer.step()
+        # 软更新
+        for target_param, param in zip(self.target_critic.parameters(), self.critic.parameters()):
+            target_param.data.copy_(
+                target_param.data * (1.0 - self.soft_tau) +
+                param.data * self.soft_tau
+            )
+        for target_param, param in zip(self.target_actor.parameters(), self.actor.parameters()):
+            target_param.data.copy_(
+                target_param.data * (1.0 - self.soft_tau) +
+                param.data * self.soft_tau
+            )
+    def save(self,path):
+        torch.save(self.actor.state_dict(), path+'checkpoint.pt')
+
+    def load(self,path):
+        self.actor.load_state_dict(torch.load(path+'checkpoint.pt')) 

projects/codes/DDPG/env.py

@@ -0,0 +1,56 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+@Author: John
+@Email: johnjim0816@gmail.com
+@Date: 2020-06-10 15:28:30
+@LastEditor: John
+LastEditTime: 2021-09-16 00:52:30
+@Discription: 
+@Environment: python 3.7.7
+'''
+import gym
+import numpy as np
+
+class NormalizedActions(gym.ActionWrapper):
+    ''' 将action范围重定在[0.1]之间
+    '''
+    def action(self, action):
+        low_bound   = self.action_space.low
+        upper_bound = self.action_space.high
+        action = low_bound + (action + 1.0) * 0.5 * (upper_bound - low_bound)
+        action = np.clip(action, low_bound, upper_bound)
+        return action
+
+    def reverse_action(self, action):
+        low_bound   = self.action_space.low
+        upper_bound = self.action_space.high
+        action = 2 * (action - low_bound) / (upper_bound - low_bound) - 1
+        action = np.clip(action, low_bound, upper_bound)
+        return action
+
+class OUNoise(object):
+    '''Ornstein–Uhlenbeck噪声
+    '''
+    def __init__(self, action_space, mu=0.0, theta=0.15, max_sigma=0.3, min_sigma=0.3, decay_period=100000):
+        self.mu           = mu # OU噪声的参数
+        self.theta        = theta # OU噪声的参数
+        self.sigma        = max_sigma # OU噪声的参数
+        self.max_sigma    = max_sigma
+        self.min_sigma    = min_sigma
+        self.decay_period = decay_period
+        self.n_actions   = action_space.shape[0]
+        self.low          = action_space.low
+        self.high         = action_space.high
+        self.reset()
+    def reset(self):
+        self.obs = np.ones(self.n_actions) * self.mu
+    def evolve_obs(self):
+        x  = self.obs
+        dx = self.theta * (self.mu - x) + self.sigma * np.random.randn(self.n_actions)
+        self.obs = x + dx
+        return self.obs
+    def get_action(self, action, t=0):
+        ou_obs = self.evolve_obs()
+        self.sigma = self.max_sigma - (self.max_sigma - self.min_sigma) * min(1.0, t / self.decay_period) # sigma会逐渐衰减
+        return np.clip(action + ou_obs, self.low, self.high) # 动作加上噪声后进行剪切

projects/codes/DDPG/outputs/Pendulum-v1/20220713-225402/results/params.json

@@ -0,0 +1,18 @@
+{
+    "algo_name": "DDPG",
+    "env_name": "Pendulum-v1",
+    "train_eps": 300,
+    "test_eps": 20,
+    "gamma": 0.99,
+    "critic_lr": 0.001,
+    "actor_lr": 0.0001,
+    "memory_capacity": 8000,
+    "batch_size": 128,
+    "target_update": 2,
+    "soft_tau": 0.01,
+    "hidden_dim": 256,
+    "deivce": "cpu",
+    "result_path": "C:\\Users\\24438\\Desktop\\rl-tutorials/outputs/DDPG/outputs/Pendulum-v1/20220713-225402/results//",
+    "model_path": "C:\\Users\\24438\\Desktop\\rl-tutorials/outputs/DDPG/outputs/Pendulum-v1/20220713-225402/models/",
+    "save_fig": true
+}

projects/codes/DDPG/task0.py

@@ -0,0 +1,133 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+@Author: John
+@Email: johnjim0816@gmail.com
+@Date: 2020-06-11 20:58:21
+@LastEditor: John
+LastEditTime: 2022-07-21 21:51:34
+@Discription: 
+@Environment: python 3.7.7
+'''
+import sys,os
+curr_path = os.path.dirname(os.path.abspath(__file__))  # current path
+parent_path = os.path.dirname(curr_path)  # parent path
+sys.path.append(parent_path)  # add to system path
+
+import datetime
+import gym
+import torch
+import argparse
+
+from env import NormalizedActions,OUNoise
+from ddpg import DDPG
+from common.utils import save_results,make_dir
+from common.utils import plot_rewards,save_args
+
+def get_args():
+    """ Hyperparameters
+    """
+    curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")  # Obtain current time
+    parser = argparse.ArgumentParser(description="hyperparameters")      
+    parser.add_argument('--algo_name',default='DDPG',type=str,help="name of algorithm")
+    parser.add_argument('--env_name',default='Pendulum-v1',type=str,help="name of environment")
+    parser.add_argument('--train_eps',default=300,type=int,help="episodes of training")
+    parser.add_argument('--test_eps',default=20,type=int,help="episodes of testing")
+    parser.add_argument('--gamma',default=0.99,type=float,help="discounted factor")
+    parser.add_argument('--critic_lr',default=1e-3,type=float,help="learning rate of critic")
+    parser.add_argument('--actor_lr',default=1e-4,type=float,help="learning rate of actor")
+    parser.add_argument('--memory_capacity',default=8000,type=int,help="memory capacity")
+    parser.add_argument('--batch_size',default=128,type=int)
+    parser.add_argument('--target_update',default=2,type=int)
+    parser.add_argument('--soft_tau',default=1e-2,type=float)
+    parser.add_argument('--hidden_dim',default=256,type=int)
+    parser.add_argument('--device',default='cpu',type=str,help="cpu or cuda") 
+    parser.add_argument('--result_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/results/' )
+    parser.add_argument('--model_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/models/' ) # path to save models
+    parser.add_argument('--save_fig',default=True,type=bool,help="if save figure or not")   
+    args = parser.parse_args()                           
+    return args
+
+def env_agent_config(cfg,seed=1):
+    env = NormalizedActions(gym.make(cfg.env_name)) # 装饰action噪声
+    env.seed(seed) # 随机种子
+    n_states = env.observation_space.shape[0]
+    n_actions = env.action_space.shape[0]
+    agent = DDPG(n_states,n_actions,cfg)
+    return env,agent
+def train(cfg, env, agent):
+    print('Start training!')
+    print(f'Env:{cfg.env_name}, Algorithm:{cfg.algo_name}, Device:{cfg.device}')
+    ou_noise = OUNoise(env.action_space)  # noise of action
+    rewards = [] # 记录所有回合的奖励
+    ma_rewards = []  # 记录所有回合的滑动平均奖励
+    for i_ep in range(cfg.train_eps):
+        state = env.reset()
+        ou_noise.reset()
+        done = False
+        ep_reward = 0
+        i_step = 0
+        while not done:
+            i_step += 1
+            action = agent.choose_action(state)
+            action = ou_noise.get_action(action, i_step) 
+            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 (i_ep+1)%10 == 0:
+            print(f'Env:{i_ep+1}/{cfg.train_eps}, Reward:{ep_reward:.2f}')
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
+        else:
+            ma_rewards.append(ep_reward)
+    print('Finish training!')
+    return {'rewards':rewards,'ma_rewards':ma_rewards}
+
+def test(cfg, env, agent):
+    print('Start testing')
+    print(f'Env:{cfg.env_name}, Algorithm:{cfg.algo_name}, Device:{cfg.device}')
+    rewards = [] # 记录所有回合的奖励
+    ma_rewards = []  # 记录所有回合的滑动平均奖励
+    for i_ep in range(cfg.test_eps):
+        state = env.reset() 
+        done = False
+        ep_reward = 0
+        i_step = 0
+        while not done:
+            i_step += 1
+            action = agent.choose_action(state)
+            next_state, reward, done, _ = env.step(action)
+            ep_reward += reward
+            state = next_state
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(0.9*ma_rewards[-1]+0.1*ep_reward)
+        else:
+            ma_rewards.append(ep_reward)
+        print(f"Epside:{i_ep+1}/{cfg.test_eps}, Reward:{ep_reward:.1f}")
+    print('Finish testing!')
+    return {'rewards':rewards,'ma_rewards':ma_rewards}
+if __name__ == "__main__":
+    cfg = get_args()
+    # training
+    env,agent = env_agent_config(cfg,seed=1)
+    res_dic = train(cfg, env, agent)
+    make_dir(cfg.result_path, cfg.model_path)
+    save_args(cfg)
+    agent.save(path=cfg.model_path)
+    save_results(res_dic, tag='train',
+                 path=cfg.result_path)  
+    plot_rewards(res_dic['rewards'], res_dic['ma_rewards'], cfg, tag="train")  
+    # testing
+    env,agent = env_agent_config(cfg,seed=10)
+    agent.load(path=cfg.model_path)
+    res_dic = test(cfg,env,agent)
+    save_results(res_dic, tag='test',
+                 path=cfg.result_path)  
+    plot_rewards(res_dic['rewards'], res_dic['ma_rewards'], cfg, tag="test")   
+

projects/codes/DQN/README.md

@@ -0,0 +1,218 @@
+# DQN
+
+## 原理简介
+
+DQN是Q-leanning算法的优化和延伸，Q-leaning中使用有限的Q表存储值的信息，而DQN中则用神经网络替代Q表存储信息，这样更适用于高维的情况，相关知识基础可参考[datawhale李宏毅笔记-Q学习](https://datawhalechina.github.io/easy-rl/#/chapter6/chapter6)。
+
+论文方面主要可以参考两篇，一篇就是2013年谷歌DeepMind团队的[Playing Atari with Deep Reinforcement Learning](https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf)，一篇是也是他们团队后来在Nature杂志上发表的[Human-level control through deep reinforcement learning](https://web.stanford.edu/class/psych209/Readings/MnihEtAlHassibis15NatureControlDeepRL.pdf)。后者在算法层面增加target q-net，也可以叫做Nature DQN。
+
+Nature DQN使用了两个Q网络，一个当前Q网络𝑄用来选择动作，更新模型参数，另一个目标Q网络𝑄′用于计算目标Q值。目标Q网络的网络参数不需要迭代更新，而是每隔一段时间从当前Q网络𝑄复制过来，即延时更新，这样可以减少目标Q值和当前的Q值相关性。
+
+要注意的是，两个Q网络的结构是一模一样的。这样才可以复制网络参数。Nature DQN和[Playing Atari with Deep Reinforcement Learning](https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf)相比，除了用一个新的相同结构的目标Q网络来计算目标Q值以外，其余部分基本是完全相同的。细节也可参考[强化学习（九）Deep Q-Learning进阶之Nature DQN](https://www.cnblogs.com/pinard/p/9756075.html)。
+
+https://blog.csdn.net/JohnJim0/article/details/109557173)
+
+## 伪代码
+
+<img src="assets/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L0pvaG5KaW0w,size_16,color_FFFFFF,t_70.png" alt="img" style="zoom:50%;" />
+
+## 代码实现
+
+### RL接口
+
+首先是强化学习训练的基本接口，即通用的训练模式：
+```python
+for i_episode in range(MAX_EPISODES):
+	state = env.reset() # reset环境状态
+	for i_step in range(MAX_STEPS):
+    action = agent.choose_action(state) # 根据当前环境state选择action
+    next_state, reward, done, _ = env.step(action) # 更新环境参数
+    agent.memory.push(state, action, reward, next_state, done) # 将state等这些transition存入memory
+    agent.update() # 每步更新网络
+    state = next_state # 跳转到下一个状态
+    if done:
+    	break        
+```
+每个episode加一个MAX_STEPS，也可以使用while not done, 加这个max_steps有时是因为比如gym环境训练目标就是在200个step下达到200的reward，或者是当完成一个episode的步数较多时也可以设置，基本流程跟所有伪代码一致，如下：
+1. agent选择动作
+2. 环境根据agent的动作反馈出next_state和reward
+3. agent进行更新，如有memory就会将transition(包含state，reward，action等)存入memory中
+4. 跳转到下一个状态
+5. 如果done了，就跳出循环，进行下一个episode的训练。
+
+想要实现完整的算法还需要创建Qnet，Replaybuffer等类
+
+### 两个Q网络
+
+上文讲了Nature DQN中有两个Q网络，一个是policy_net，一个是延时更新的target_net，两个网络的结构是一模一样的，如下(见```model.py```)，注意DQN使用的Qnet就是全连接网络即FCH：
+```python
+import torch.nn as nn
+import torch.nn.functional as F
+
+class FCN(nn.Module):
+    def __init__(self, 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(FCN, 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)
+```
+输入为n_states，输出为n_actions，包含一个128维度的隐藏层，这里根据需要可增加隐藏层维度和数量，然后一般使用relu激活函数，这里跟深度学习的网路设置是一样的。
+
+### Replay Buffer
+
+然后就是Replay Memory了，其作用主要是是克服经验数据的相关性（correlated data）和非平稳分布（non-stationary distribution）问题，实现如下(见```memory.py```)：
+
+```python
+import random
+import numpy as np
+
+class ReplayBuffer:
+    
+    def __init__(self, capacity):
+        self.capacity = capacity
+        self.buffer = []
+        self.position = 0
+    
+    def push(self, state, action, reward, next_state, done):
+        if len(self.buffer) < self.capacity:
+            self.buffer.append(None)
+        self.buffer[self.position] = (state, action, reward, next_state, done)
+        self.position = (self.position + 1) % self.capacity
+    
+    def sample(self, batch_size):
+        batch = random.sample(self.buffer, batch_size)
+        state, action, reward, next_state, done =  zip(*batch)
+        return state, action, reward, next_state, done
+    
+    def __len__(self):
+        return len(self.buffer)
+```
+
+参数capacity表示buffer的容量，主要包括push和sample两个步骤，push是将transitions放到memory中，sample是从memory随机抽取一些transition。
+
+### Agent类
+
+在```agent.py```中我们定义强化学习算法类，包括```choose_action```(选择动作，使用e-greedy策略时会多一个```predict```函数，下面会将到)和```update```(更新)等函数。
+
+在类中建立两个网络，以及optimizer和memory，
+
+```python
+self.policy_net = MLP(n_states, n_actions,hidden_dim=cfg.hidden_dim).to(self.device)
+self.target_net = MLP(n_states, n_actions,hidden_dim=cfg.hidden_dim).to(self.device)
+for target_param, param in zip(self.target_net.parameters(),self.policy_net.parameters()): # copy params from policy net
+    target_param.data.copy_(param.data)
+self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg.lr)
+self.memory = ReplayBuffer(cfg.memory_capacity)
+```
+然后是选择action：
+
+```python
+def choose_action(self, state):
+        '''选择动作
+        '''
+    self.frame_idx += 1
+    if random.random() > self.epsilon(self.frame_idx):
+        action = self.predict(state)
+    else:
+        action = random.randrange(self.n_actions)
+    return action
+```
+
+这里使用e-greedy策略，即设置一个参数epsilon，如果生成的随机数大于epsilon，就根据网络预测的选择action，否则还是随机选择action，这个epsilon是会逐渐减小的，可以使用线性或者指数减小的方式，但不会减小到零，这样在训练稳定时还能保持一定的探索，这部分可以学习探索与利用(exploration and exploition)相关知识。
+
+上面讲到的预测函数其实就是根据state选取q值最大的action，如下：
+
+```python
+def predict(self,state):
+    with torch.no_grad():
+        state = torch.tensor([state], device=self.device, dtype=torch.float32)
+        q_values = self.policy_net(state)
+        action = q_values.max(1)[1].item()
+```
+
+然后是更新函数了：
+
+```python
+def update(self):
+
+        if len(self.memory) < self.batch_size:
+            return
+        # 从memory中随机采样transition
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(
+            self.batch_size)
+        '''转为张量
+        例如tensor([[-4.5543e-02, -2.3910e-01,  1.8344e-02,  2.3158e-01],...,[-1.8615e-02, -2.3921e-01, -1.1791e-02,  2.3400e-01]])'''
+        state_batch = torch.tensor(
+            state_batch, device=self.device, dtype=torch.float)
+        action_batch = torch.tensor(action_batch, device=self.device).unsqueeze(
+            1)  # 例如tensor([[1],...,[0]])
+        reward_batch = torch.tensor(
+            reward_batch, device=self.device, dtype=torch.float)  # tensor([1., 1.,...,1])
+        next_state_batch = torch.tensor(
+            next_state_batch, device=self.device, dtype=torch.float)
+        done_batch = torch.tensor(np.float32(
+            done_batch), device=self.device)
+
+        '''计算当前(s_t,a)对应的Q(s_t, a)'''
+        '''torch.gather:对于a=torch.Tensor([[1,2],[3,4]]),那么a.gather(1,torch.Tensor([[0],[1]]))=torch.Tensor([[1],[3]])'''
+        q_values = self.policy_net(state_batch).gather(
+            dim=1, index=action_batch)  # 等价于self.forward
+        # 计算所有next states的V(s_{t+1})，即通过target_net中选取reward最大的对应states
+        next_q_values = self.target_net(next_state_batch).max(
+            1)[0].detach()  # 比如tensor([ 0.0060, -0.0171,...,])
+        # 计算 expected_q_value
+        # 对于终止状态，此时done_batch[0]=1, 对应的expected_q_value等于reward
+        expected_q_values = reward_batch + \
+            self.gamma * next_q_values * (1-done_batch)
+        # self.loss = F.smooth_l1_loss(q_values,expected_q_values.unsqueeze(1)) # 计算 Huber loss
+        loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1))  # 计算 均方误差loss
+        # 优化模型
+        self.optimizer.zero_grad()  # zero_grad清除上一步所有旧的gradients from the last step
+        # loss.backward()使用backpropagation计算loss相对于所有parameters(需要gradients)的微分
+        loss.backward()
+        # for param in self.policy_net.parameters():  # clip防止梯度爆炸
+        #     param.grad.data.clamp_(-1, 1)
+        self.optimizer.step()  # 更新模型
+```
+
+更新遵循伪代码的以下部分：
+
+<img src="assets/image-20210507162813393.png" alt="image-20210507162813393" style="zoom:50%;" />
+
+首先从replay buffer中选取一个batch的数据，计算loss，然后进行minibatch SGD。
+
+然后是保存与加载模型的部分，如下：
+
+```python
+def save(self, path):
+        torch.save(self.target_net.state_dict(), path+'dqn_checkpoint.pth')
+def load(self, path):
+    self.target_net.load_state_dict(torch.load(path+'dqn_checkpoint.pth'))
+    for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()):
+        param.data.copy_(target_param.data)
+```
+
+
+
+### 实验结果
+
+训练结果如下：
+
+<img src="assets/train_rewards_curve.png" alt="train_rewards_curve" style="zoom: 67%;" />
+
+<img src="assets/eval_rewards_curve.png" alt="eval_rewards_curve" style="zoom:67%;" />
+
+## 参考
+
+[with torch.no_grad()](https://www.jianshu.com/p/1cea017f5d11)
+

projects/codes/DQN/dqn.py

@@ -0,0 +1,126 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+@Author: John
+@Email: johnjim0816@gmail.com
+@Date: 2020-06-12 00:50:49
+@LastEditor: John
+LastEditTime: 2022-07-20 23:57:16
+@Discription: 
+@Environment: python 3.7.7
+'''
+'''off-policy
+'''
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.optim as optim
+import random
+import math
+import numpy as np
+
+class MLP(nn.Module):
+    def __init__(self, n_states,n_actions,hidden_dim=128):
+        """ 初始化q网络，为全连接网络
+            n_states: 输入的特征数即环境的状态维度
+            n_actions: 输出的动作维度
+        """
+        super(MLP, self).__init__()
+        self.fc1 = nn.Linear(n_states, hidden_dim) # 输入层
+        self.fc2 = nn.Linear(hidden_dim,hidden_dim) # 隐藏层
+        self.fc3 = nn.Linear(hidden_dim, n_actions) # 输出层
+        
+    def forward(self, x):
+        # 各层对应的激活函数
+        x = F.relu(self.fc1(x)) 
+        x = F.relu(self.fc2(x))
+        return self.fc3(x)
+
+class ReplayBuffer:
+    def __init__(self, capacity):
+        self.capacity = capacity # 经验回放的容量
+        self.buffer = [] # 缓冲区
+        self.position = 0 
+    
+    def push(self, state, action, reward, next_state, done):
+        ''' 缓冲区是一个队列，容量超出时去掉开始存入的转移(transition)
+        '''
+        if len(self.buffer) < self.capacity:
+            self.buffer.append(None)
+        self.buffer[self.position] = (state, action, reward, next_state, done)
+        self.position = (self.position + 1) % self.capacity 
+    
+    def sample(self, batch_size):
+        batch = random.sample(self.buffer, batch_size) # 随机采出小批量转移
+        state, action, reward, next_state, done =  zip(*batch) # 解压成状态，动作等
+        return state, action, reward, next_state, done
+    
+    def __len__(self):
+        ''' 返回当前存储的量
+        '''
+        return len(self.buffer)
+
+class DQN:
+    def __init__(self, n_states,n_actions,cfg):
+
+        self.n_actions = n_actions  
+        self.device = torch.device(cfg.device)  # cpu or cuda
+        self.gamma = cfg.gamma  # 奖励的折扣因子
+        # e-greedy策略相关参数
+        self.frame_idx = 0  # 用于epsilon的衰减计数
+        self.epsilon = lambda frame_idx: cfg.epsilon_end + \
+            (cfg.epsilon_start - cfg.epsilon_end) * \
+            math.exp(-1. * frame_idx / cfg.epsilon_decay)
+        self.batch_size = cfg.batch_size
+        self.policy_net = MLP(n_states,n_actions).to(self.device)
+        self.target_net = MLP(n_states,n_actions).to(self.device)
+        for target_param, param in zip(self.target_net.parameters(),self.policy_net.parameters()): # 复制参数到目标网路targe_net
+            target_param.data.copy_(param.data)
+        self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg.lr) # 优化器
+        self.memory = ReplayBuffer(cfg.memory_capacity) # 经验回放
+
+    def choose_action(self, state):
+        ''' 选择动作
+        '''
+        self.frame_idx += 1
+        if random.random() > self.epsilon(self.frame_idx):
+            with torch.no_grad():
+                state = torch.tensor(state, device=self.device, dtype=torch.float32).unsqueeze(dim=0)
+                q_values = self.policy_net(state)
+                action = q_values.max(1)[1].item() # 选择Q值最大的动作
+        else:
+            action = random.randrange(self.n_actions)
+        return action
+    def update(self):
+        if len(self.memory) < self.batch_size: # 当memory中不满足一个批量时，不更新策略
+            return
+        # 从经验回放中(replay memory)中随机采样一个批量的转移(transition)
+        # print('updating')
+        
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(
+            self.batch_size)
+        state_batch = torch.tensor(np.array(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(np.array(next_state_batch), device=self.device, dtype=torch.float)
+        done_batch = torch.tensor(np.float32(done_batch), device=self.device)
+        q_values = self.policy_net(state_batch).gather(dim=1, index=action_batch) # 计算当前状态(s_t,a)对应的Q(s_t, a)
+        next_q_values = self.target_net(next_state_batch).max(1)[0].detach() # 计算下一时刻的状态(s_t_,a)对应的Q值
+        # 计算期望的Q值，对于终止状态，此时done_batch[0]=1, 对应的expected_q_value等于reward
+        expected_q_values = reward_batch + self.gamma * next_q_values * (1-done_batch)
+        loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1))  # 计算均方根损失
+        # 优化更新模型
+        self.optimizer.zero_grad()  
+        loss.backward()
+        for param in self.policy_net.parameters():  # clip防止梯度爆炸
+            param.grad.data.clamp_(-1, 1)
+        self.optimizer.step() 
+
+    def save(self, path):
+        torch.save(self.target_net.state_dict(), path+'dqn_checkpoint.pth')
+
+    def load(self, path):
+        self.target_net.load_state_dict(torch.load(path+'dqn_checkpoint.pth'))
+        for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()):
+            param.data.copy_(target_param.data)

projects/codes/DQN/dqn_cnn.py

@@ -0,0 +1,134 @@
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import torch.autograd as autograd 
+import random
+import math
+class CNN(nn.Module):
+    def __init__(self, input_dim, output_dim):
+        super(CNN, self).__init__()
+        
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        
+        self.features = nn.Sequential(
+            nn.Conv2d(input_dim[0], 32, kernel_size=8, stride=4),
+            nn.ReLU(),
+            nn.Conv2d(32, 64, kernel_size=4, stride=2),
+            nn.ReLU(),
+            nn.Conv2d(64, 64, kernel_size=3, stride=1),
+            nn.ReLU()
+        )
+        
+        self.fc = nn.Sequential(
+            nn.Linear(self.feature_size(), 512),
+            nn.ReLU(),
+            nn.Linear(512, self.output_dim)
+        )
+        
+    def forward(self, x):
+        x = self.features(x)
+        x = x.view(x.size(0), -1)
+        x = self.fc(x)
+        return x
+    
+    def feature_size(self):
+        return self.features(autograd.Variable(torch.zeros(1, *self.input_dim))).view(1, -1).size(1)
+
+
+    def act(self, state, epsilon):
+        if random.random() > epsilon:
+            state   = Variable(torch.FloatTensor(np.float32(state)).unsqueeze(0), volatile=True)
+            q_value = self.forward(state)
+            action  = q_value.max(1)[1].data[0]
+        else:
+            action = random.randrange(env.action_space.n)
+        return action
+
+class ReplayBuffer:
+    def __init__(self, capacity):
+        self.capacity = capacity # 经验回放的容量
+        self.buffer = [] # 缓冲区
+        self.position = 0 
+    
+    def push(self, state, action, reward, next_state, done):
+        ''' 缓冲区是一个队列，容量超出时去掉开始存入的转移(transition)
+        '''
+        if len(self.buffer) < self.capacity:
+            self.buffer.append(None)
+        self.buffer[self.position] = (state, action, reward, next_state, done)
+        self.position = (self.position + 1) % self.capacity 
+    
+    def sample(self, batch_size):
+        batch = random.sample(self.buffer, batch_size) # 随机采出小批量转移
+        state, action, reward, next_state, done =  zip(*batch) # 解压成状态，动作等
+        return state, action, reward, next_state, done
+    
+    def __len__(self):
+        ''' 返回当前存储的量
+        '''
+        return len(self.buffer)
+
+class DQN:
+    def __init__(self, n_states, n_actions, cfg):
+
+        self.n_actions = n_actions  # 总的动作个数
+        self.device = cfg.device  # 设备，cpu或gpu等
+        self.gamma = cfg.gamma  # 奖励的折扣因子
+        # e-greedy策略相关参数
+        self.frame_idx = 0  # 用于epsilon的衰减计数
+        self.epsilon = lambda frame_idx: cfg.epsilon_end + \
+            (cfg.epsilon_start - cfg.epsilon_end) * \
+            math.exp(-1. * frame_idx / cfg.epsilon_decay)
+        self.batch_size = cfg.batch_size
+        self.policy_net = CNN(n_states, n_actions).to(self.device)
+        self.target_net = CNN(n_states, n_actions).to(self.device)
+        for target_param, param in zip(self.target_net.parameters(),self.policy_net.parameters()): # 复制参数到目标网路targe_net
+            target_param.data.copy_(param.data)
+        self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg.lr) # 优化器
+        self.memory = ReplayBuffer(cfg.memory_capacity) # 经验回放
+
+    def choose_action(self, state):
+        ''' 选择动作
+        '''
+        self.frame_idx += 1
+        if random.random() > self.epsilon(self.frame_idx):
+            with torch.no_grad():
+                print(type(state))
+                state = torch.tensor([state], device=self.device, dtype=torch.float32)
+                q_values = self.policy_net(state)
+                action = q_values.max(1)[1].item() # 选择Q值最大的动作
+        else:
+            action = random.randrange(self.n_actions)
+        return action
+    def update(self):
+        if len(self.memory) < self.batch_size: # 当memory中不满足一个批量时，不更新策略
+            return
+        # 从经验回放中(replay memory)中随机采样一个批量的转移(transition)
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(
+            self.batch_size)
+        # 转为张量
+        state_batch = torch.tensor(state_batch, device=self.device, dtype=torch.float)
+        action_batch = torch.tensor(action_batch, device=self.device).unsqueeze(1)  
+        reward_batch = torch.tensor(reward_batch, device=self.device, dtype=torch.float)  
+        next_state_batch = torch.tensor(next_state_batch, device=self.device, dtype=torch.float)
+        done_batch = torch.tensor(np.float32(done_batch), device=self.device)
+        q_values = self.policy_net(state_batch).gather(dim=1, index=action_batch) # 计算当前状态(s_t,a)对应的Q(s_t, a)
+        next_q_values = self.target_net(next_state_batch).max(1)[0].detach() # 计算下一时刻的状态(s_t_,a)对应的Q值
+        # 计算期望的Q值，对于终止状态，此时done_batch[0]=1, 对应的expected_q_value等于reward
+        expected_q_values = reward_batch + self.gamma * next_q_values * (1-done_batch)
+        loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1))  # 计算均方根损失
+        # 优化更新模型
+        self.optimizer.zero_grad()  
+        loss.backward()
+        for param in self.policy_net.parameters():  # clip防止梯度爆炸
+            param.grad.data.clamp_(-1, 1)
+        self.optimizer.step() 
+
+    def save(self, path):
+        torch.save(self.target_net.state_dict(), path+'dqn_checkpoint.pth')
+
+    def load(self, path):
+        self.target_net.load_state_dict(torch.load(path+'dqn_checkpoint.pth'))
+        for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()):
+            param.data.copy_(target_param.data)

projects/codes/DQN/dqn_cnn2.py

@@ -0,0 +1,142 @@
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import torch.autograd as autograd 
+import random
+import math
+import numpy as np
+class CNN(nn.Module):
+    def __init__(self, n_frames, n_actions):
+        super(CNN,self).__init__()
+        self.n_frames = n_frames
+        self.n_actions = n_actions
+        
+        # Layers
+        self.conv1 = nn.Conv2d(
+            in_channels=n_frames,
+            out_channels=16,
+            kernel_size=8,
+            stride=4,
+            padding=2
+            )
+        self.conv2 = nn.Conv2d(
+            in_channels=16,
+            out_channels=32,
+            kernel_size=4,
+            stride=2,
+            padding=1
+            )
+        self.fc1 = nn.Linear(
+            in_features=3200,
+            out_features=256,
+            )
+        self.fc2 = nn.Linear(
+            in_features=256,
+            out_features=n_actions,
+            )
+        
+        # Activation Functions
+        self.relu = nn.ReLU()
+    
+    def flatten(self, x):
+        batch_size = x.size()[0]
+        x = x.view(batch_size, -1)
+        return x
+    
+    def forward(self, x):
+        
+        # Forward pass
+        x = self.relu(self.conv1(x))  # In: (80, 80, 4)  Out: (20, 20, 16)
+        x = self.relu(self.conv2(x))  # In: (20, 20, 16) Out: (10, 10, 32)
+        x = self.flatten(x)           # In: (10, 10, 32) Out: (3200,)
+        x = self.relu(self.fc1(x))    # In: (3200,)      Out: (256,)
+        x = self.fc2(x)               # In: (256,)       Out: (4,)
+        
+        return x
+
+class ReplayBuffer:
+    def __init__(self, capacity):
+        self.capacity = capacity # 经验回放的容量
+        self.buffer = [] # 缓冲区
+        self.position = 0 
+    
+    def push(self, state, action, reward, next_state, done):
+        ''' 缓冲区是一个队列，容量超出时去掉开始存入的转移(transition)
+        '''
+        if len(self.buffer) < self.capacity:
+            self.buffer.append(None)
+        self.buffer[self.position] = (state, action, reward, next_state, done)
+        self.position = (self.position + 1) % self.capacity 
+    
+    def sample(self, batch_size):
+        batch = random.sample(self.buffer, batch_size) # 随机采出小批量转移
+        state, action, reward, next_state, done =  zip(*batch) # 解压成状态，动作等
+        return state, action, reward, next_state, done
+    
+    def __len__(self):
+        ''' 返回当前存储的量
+        '''
+        return len(self.buffer)
+
+class DQN:
+    def __init__(self, n_states, n_actions, cfg):
+
+        self.n_actions = n_actions  # 总的动作个数
+        self.device = cfg.device  # 设备，cpu或gpu等
+        self.gamma = cfg.gamma  # 奖励的折扣因子
+        # e-greedy策略相关参数
+        self.frame_idx = 0  # 用于epsilon的衰减计数
+        self.epsilon = lambda frame_idx: cfg.epsilon_end + \
+            (cfg.epsilon_start - cfg.epsilon_end) * \
+            math.exp(-1. * frame_idx / cfg.epsilon_decay)
+        self.batch_size = cfg.batch_size
+        self.policy_net = CNN(n_states, n_actions).to(self.device)
+        self.target_net = CNN(n_states, n_actions).to(self.device)
+        for target_param, param in zip(self.target_net.parameters(),self.policy_net.parameters()): # 复制参数到目标网路targe_net
+            target_param.data.copy_(param.data)
+        self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg.lr) # 优化器
+        self.memory = ReplayBuffer(cfg.memory_capacity) # 经验回放
+
+    def choose_action(self, state):
+        ''' 选择动作
+        '''
+        self.frame_idx += 1
+        if random.random() > self.epsilon(self.frame_idx):
+            with torch.no_grad():
+                state = torch.tensor([state], device=self.device, dtype=torch.float32)
+                q_values = self.policy_net(state)
+                action = q_values.max(1)[1].item() # 选择Q值最大的动作
+        else:
+            action = random.randrange(self.n_actions)
+        return action
+    def update(self):
+        if len(self.memory) < self.batch_size: # 当memory中不满足一个批量时，不更新策略
+            return
+        # 从经验回放中(replay memory)中随机采样一个批量的转移(transition)
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(
+            self.batch_size)
+        # 转为张量
+        state_batch = torch.tensor(state_batch, device=self.device, dtype=torch.float)
+        action_batch = torch.tensor(action_batch, device=self.device).unsqueeze(1)  
+        reward_batch = torch.tensor(reward_batch, device=self.device, dtype=torch.float)  
+        next_state_batch = torch.tensor(next_state_batch, device=self.device, dtype=torch.float)
+        done_batch = torch.tensor(np.float32(done_batch), device=self.device)
+        q_values = self.policy_net(state_batch).gather(dim=1, index=action_batch) # 计算当前状态(s_t,a)对应的Q(s_t, a)
+        next_q_values = self.target_net(next_state_batch).max(1)[0].detach() # 计算下一时刻的状态(s_t_,a)对应的Q值
+        # 计算期望的Q值，对于终止状态，此时done_batch[0]=1, 对应的expected_q_value等于reward
+        expected_q_values = reward_batch + self.gamma * next_q_values * (1-done_batch)
+        loss = nn.MSELoss()(q_values, expected_q_values.unsqueeze(1))  # 计算均方根损失
+        # 优化更新模型
+        self.optimizer.zero_grad()  
+        loss.backward()
+        for param in self.policy_net.parameters():  # clip防止梯度爆炸
+            param.grad.data.clamp_(-1, 1)
+        self.optimizer.step() 
+
+    def save(self, path):
+        torch.save(self.target_net.state_dict(), path+'dqn_checkpoint.pth')
+
+    def load(self, path):
+        self.target_net.load_state_dict(torch.load(path+'dqn_checkpoint.pth'))
+        for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()):
+            param.data.copy_(target_param.data)

projects/codes/DQN/outputs/CartPole-v0/20220713-211653/results/params.json

@@ -0,0 +1,19 @@
+{
+    "algo_name": "DQN",
+    "env_name": "CartPole-v0",
+    "train_eps": 200,
+    "test_eps": 20,
+    "gamma": 0.95,
+    "epsilon_start": 0.95,
+    "epsilon_end": 0.01,
+    "epsilon_decay": 500,
+    "lr": 0.0001,
+    "memory_capacity": 100000,
+    "batch_size": 64,
+    "target_update": 4,
+    "hidden_dim": 256,
+    "deivce": "cpu",
+    "result_path": "C:\\Users\\24438\\Desktop\\rl-tutorials/outputs/CartPole-v0/20220713-211653/results/",
+    "model_path": "C:\\Users\\24438\\Desktop\\rl-tutorials/outputs/CartPole-v0/20220713-211653/models/",
+    "save_fig": true
+}

projects/codes/DQN/task0.py

@@ -0,0 +1,148 @@
+import sys,os
+curr_path = os.path.dirname(os.path.abspath(__file__))  # current path
+parent_path = os.path.dirname(curr_path)  # parent path
+sys.path.append(parent_path)  # add to system path
+import torch.nn as nn
+import torch.nn.functional as F
+
+import gym
+import torch
+import datetime
+import numpy as np
+import argparse
+from common.utils import save_results, make_dir
+from common.utils import plot_rewards,save_args
+from dqn import DQN
+
+def get_args():
+    """ Hyperparameters
+    """
+    curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")  # Obtain current time
+    parser = argparse.ArgumentParser(description="hyperparameters")      
+    parser.add_argument('--algo_name',default='DQN',type=str,help="name of algorithm")
+    parser.add_argument('--env_name',default='CartPole-v0',type=str,help="name of environment")
+    parser.add_argument('--train_eps',default=200,type=int,help="episodes of training")
+    parser.add_argument('--test_eps',default=20,type=int,help="episodes of testing")
+    parser.add_argument('--gamma',default=0.95,type=float,help="discounted factor")
+    parser.add_argument('--epsilon_start',default=0.95,type=float,help="initial value of epsilon")
+    parser.add_argument('--epsilon_end',default=0.01,type=float,help="final value of epsilon")
+    parser.add_argument('--epsilon_decay',default=500,type=int,help="decay rate of epsilon")
+    parser.add_argument('--lr',default=0.0001,type=float,help="learning rate")
+    parser.add_argument('--memory_capacity',default=100000,type=int,help="memory capacity")
+    parser.add_argument('--batch_size',default=64,type=int)
+    parser.add_argument('--target_update',default=4,type=int)
+    parser.add_argument('--hidden_dim',default=256,type=int)
+    parser.add_argument('--device',default='cpu',type=str,help="cpu or cuda") 
+    parser.add_argument('--result_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/results/' )
+    parser.add_argument('--model_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/models/' ) # path to save models
+    parser.add_argument('--save_fig',default=True,type=bool,help="if save figure or not")           
+    args = parser.parse_args()                          
+    return args
+
+def env_agent_config(cfg,seed=1):
+    ''' 创建环境和智能体
+    '''
+    env = gym.make(cfg.env_name)  # 创建环境
+    n_states = env.observation_space.shape[0]  # 状态维度
+    n_actions = env.action_space.n  # 动作维度
+    print(f"n states: {n_states}, n actions: {n_actions}")
+    agent = DQN(n_states,n_actions, cfg)  # 创建智能体
+    if seed !=0: # 设置随机种子
+        torch.manual_seed(seed)
+        env.seed(seed)
+        np.random.seed(seed)
+    return env, agent
+
+def train(cfg, env, agent):
+    ''' Training
+    '''
+    print('Start training!')
+    print(f'Env:{cfg.env_name}, A{cfg.algo_name}, 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
+    steps = []
+    for i_ep in range(cfg.train_eps):
+        ep_reward = 0  # 记录一回合内的奖励
+        ep_step = 0
+        state = env.reset()  # 重置环境，返回初始状态
+        while True:
+            ep_step += 1
+            action = agent.choose_action(state)  # 选择动作
+            next_state, reward, done, _ = env.step(action)  # 更新环境，返回transition
+            agent.memory.push(state, action, reward,
+                              next_state, done)  # 保存transition
+            state = next_state  # 更新下一个状态
+            agent.update()  # 更新智能体
+            ep_reward += reward  # 累加奖励
+            if done:
+                break
+        if (i_ep + 1) % cfg.target_update == 0:  # 智能体目标网络更新
+            agent.target_net.load_state_dict(agent.policy_net.state_dict())
+        steps.append(ep_step)
+        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) % 1 == 0:
+            print(f'Episode：{i_ep+1}/{cfg.train_eps}, Reward:{ep_reward:.2f}, Step:{ep_step:.2f} Epislon:{agent.epsilon(agent.frame_idx):.3f}')
+    print('Finish training!')
+    env.close()
+    res_dic = {'rewards':rewards,'ma_rewards':ma_rewards,'steps':steps}
+    return res_dic
+
+
+def test(cfg, env, agent):
+    print('Start testing!')
+    print(f'Env:{cfg.env_name}, A{cfg.algo_name}, 设备：{cfg.device}')
+    ############# 由于测试不需要使用epsilon-greedy策略，所以相应的值设置为0 ###############
+    cfg.epsilon_start = 0.0  # e-greedy策略中初始epsilon
+    cfg.epsilon_end = 0.0  # e-greedy策略中的终止epsilon
+    ################################################################################
+    rewards = []  # 记录所有回合的奖励
+    ma_rewards = []  # 记录所有回合的滑动平均奖励
+    steps = []
+    for i_ep in range(cfg.test_eps):
+        ep_reward = 0  # 记录一回合内的奖励
+        ep_step = 0
+        state = env.reset()  # 重置环境，返回初始状态
+        while True:
+            ep_step+=1
+            action = agent.choose_action(state)  # 选择动作
+            next_state, reward, done, _ = env.step(action)  # 更新环境，返回transition
+            state = next_state  # 更新下一个状态
+            ep_reward += reward  # 累加奖励
+            if done:
+                break
+        steps.append(ep_step)
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(ma_rewards[-1] * 0.9 + ep_reward * 0.1)
+        else:
+            ma_rewards.append(ep_reward)
+        print(f'Episode：{i_ep+1}/{cfg.test_eps}, Reward:{ep_reward:.2f}, Step:{ep_step:.2f}')
+    print('Finish testing')
+    env.close()
+    return {'rewards':rewards,'ma_rewards':ma_rewards,'steps':steps}
+
+
+if __name__ == "__main__":
+    cfg = get_args()
+    # 训练
+    env, agent = env_agent_config(cfg)
+    res_dic = train(cfg, env, agent)
+    make_dir(cfg.result_path, cfg.model_path)  
+    save_args(cfg) # save parameters
+    agent.save(path=cfg.model_path)  # save model
+    save_results(res_dic, tag='train',
+                 path=cfg.result_path)  
+    plot_rewards(res_dic['rewards'], res_dic['ma_rewards'], cfg, tag="train")  
+    # 测试
+    env, agent = env_agent_config(cfg)
+    agent.load(path=cfg.model_path)  # 导入模型
+    res_dic = test(cfg, env, agent)
+    save_results(res_dic, tag='test',
+                 path=cfg.result_path)  # 保存结果
+    plot_rewards(res_dic['rewards'], res_dic['ma_rewards'],cfg, tag="test")  # 画出结果

projects/codes/DoubleDQN/README.md

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

projects/codes/DoubleDQN/double_dqn.py

@@ -0,0 +1,160 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+@Author: John
+@Email: johnjim0816@gmail.com
+@Date: 2020-06-12 00:50:49
+@LastEditor: John
+LastEditTime: 2022-07-21 00:08:26
+@Discription: 
+@Environment: python 3.7.7
+'''
+'''off-policy
+'''
+
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import torch.nn.functional as F
+import random
+import math
+import numpy as np
+
+class ReplayBuffer:
+    def __init__(self, capacity):
+        self.capacity = capacity # 经验回放的容量
+        self.buffer = [] # 缓冲区
+        self.position = 0 
+    
+    def push(self, state, action, reward, next_state, done):
+        ''' 缓冲区是一个队列，容量超出时去掉开始存入的转移(transition)
+        '''
+        if len(self.buffer) < self.capacity:
+            self.buffer.append(None)
+        self.buffer[self.position] = (state, action, reward, next_state, done)
+        self.position = (self.position + 1) % self.capacity 
+    
+    def sample(self, batch_size):
+        batch = random.sample(self.buffer, batch_size) # 随机采出小批量转移
+        state, action, reward, next_state, done =  zip(*batch) # 解压成状态，动作等
+        return state, action, reward, next_state, done
+    
+    def __len__(self):
+        ''' 返回当前存储的量
+        '''
+        return len(self.buffer)
+
+class MLP(nn.Module):
+    def __init__(self, n_states,n_actions,hidden_dim=128):
+        """ 初始化q网络，为全连接网络
+            n_states: 输入的特征数即环境的状态维度
+            n_actions: 输出的动作维度
+        """
+        super(MLP, self).__init__()
+        self.fc1 = nn.Linear(n_states, hidden_dim) # 输入层
+        self.fc2 = nn.Linear(hidden_dim,hidden_dim) # 隐藏层
+        self.fc3 = nn.Linear(hidden_dim, n_actions) # 输出层
+        
+    def forward(self, x):
+        # 各层对应的激活函数
+        x = F.relu(self.fc1(x)) 
+        x = F.relu(self.fc2(x))
+        return self.fc3(x)
+        
+class DoubleDQN:
+    def __init__(self, n_states, n_actions, cfg):
+        self.n_actions = n_actions  # 总的动作个数
+        self.device = torch.device(cfg.device)  # 设备，cpu或gpu等
+        self.gamma = cfg.gamma
+        # e-greedy策略相关参数
+        self.actions_count = 0
+        self.epsilon_start = cfg.epsilon_start
+        self.epsilon_end = cfg.epsilon_end
+        self.epsilon_decay = cfg.epsilon_decay
+        self.batch_size = cfg.batch_size
+        self.policy_net = MLP(n_states, n_actions,hidden_dim=cfg.hidden_dim).to(self.device)
+        self.target_net = MLP(n_states, n_actions,hidden_dim=cfg.hidden_dim).to(self.device)
+        # target_net copy from policy_net
+        for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()):
+            target_param.data.copy_(param.data)
+        # self.target_net.eval()  # 不启用 BatchNormalization 和 Dropout
+        # 可查parameters()与state_dict()的区别，前者require_grad=True
+        self.optimizer = optim.Adam(self.policy_net.parameters(), lr=cfg.lr)
+        self.loss = 0
+        self.memory = ReplayBuffer(cfg.memory_capacity)
+        
+    def choose_action(self, state):
+        '''选择动作
+        '''
+        self.actions_count += 1
+        self.epsilon = self.epsilon_end + (self.epsilon_start - self.epsilon_end) * math.exp(-1. * self.actions_count / self.epsilon_decay)
+        if random.random() > self.epsilon:
+            with torch.no_grad():
+                # 先转为张量便于丢给神经网络,state元素数据原本为float64
+                # 注意state=torch.tensor(state).unsqueeze(0)跟state=torch.tensor([state])等价
+                state = torch.tensor(
+                    [state], device=self.device, dtype=torch.float32)
+                # 如tensor([[-0.0798, -0.0079]], grad_fn=<AddmmBackward>)
+                q_value = self.policy_net(state)
+                # tensor.max(1)返回每行的最大值以及对应的下标，
+                # 如torch.return_types.max(values=tensor([10.3587]),indices=tensor([0]))
+                # 所以tensor.max(1)[1]返回最大值对应的下标，即action
+                action = q_value.max(1)[1].item()  
+        else:
+            action = random.randrange(self.n_actions)
+        return action
+    def update(self):
+
+        if len(self.memory) < self.batch_size:
+            return
+        # 从memory中随机采样transition
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(
+            self.batch_size)
+        # convert to tensor
+        state_batch = torch.tensor(
+            state_batch, device=self.device, dtype=torch.float)
+        action_batch = torch.tensor(action_batch, device=self.device).unsqueeze(
+            1)  # 例如tensor([[1],...,[0]])
+        reward_batch = torch.tensor(
+            reward_batch, device=self.device, dtype=torch.float)  # tensor([1., 1.,...,1])
+        next_state_batch = torch.tensor(
+            next_state_batch, device=self.device, dtype=torch.float)
+        
+        done_batch = torch.tensor(np.float32(
+            done_batch), device=self.device)  # 将bool转为float然后转为张量
+        # 计算当前(s_t,a)对应的Q(s_t, a)
+        q_values = self.policy_net(state_batch) 
+        next_q_values = self.policy_net(next_state_batch)
+        # 代入当前选择的action，得到Q(s_t|a=a_t)
+        q_value = q_values.gather(dim=1, index=action_batch)
+        '''以下是Nature DQN的q_target计算方式
+        # 计算所有next states的Q'(s_{t+1})的最大值，Q'为目标网络的q函数
+        next_q_state_value = self.target_net(
+            next_state_batch).max(1)[0].detach()  # 比如tensor([ 0.0060, -0.0171,...,])
+        # 计算 q_target
+        # 对于终止状态，此时done_batch[0]=1, 对应的expected_q_value等于reward
+        q_target = reward_batch + self.gamma * next_q_state_value * (1-done_batch[0])
+        '''
+        '''以下是Double DQN q_target计算方式，与NatureDQN稍有不同'''
+        next_target_values = self.target_net(
+            next_state_batch)
+        # 选出Q(s_t‘, a)对应的action，代入到next_target_values获得target net对应的next_q_value，即Q’(s_t|a=argmax Q(s_t‘, a))
+        next_target_q_value = next_target_values.gather(1, torch.max(next_q_values, 1)[1].unsqueeze(1)).squeeze(1)
+        q_target = reward_batch + self.gamma * next_target_q_value * (1-done_batch)
+        self.loss = nn.MSELoss()(q_value, q_target.unsqueeze(1))  # 计算 均方误差loss
+        # 优化模型
+        self.optimizer.zero_grad()  # zero_grad清除上一步所有旧的gradients from the last step
+        # loss.backward()使用backpropagation计算loss相对于所有parameters(需要gradients)的微分
+        self.loss.backward()
+        for param in self.policy_net.parameters():  # clip防止梯度爆炸
+            param.grad.data.clamp_(-1, 1)
+        self.optimizer.step()  # 更新模型
+
+    def save(self,path):
+        torch.save(self.target_net.state_dict(), path+'checkpoint.pth')
+
+    def load(self,path):
+        self.target_net.load_state_dict(torch.load(path+'checkpoint.pth'))  
+        for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()):
+            param.data.copy_(target_param.data)  

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

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

projects/codes/DoubleDQN/task0.py

@@ -0,0 +1,138 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: JiangJi
+Email: johnjim0816@gmail.com
+Date: 2021-11-07 18:10:37
+LastEditor: JiangJi
+LastEditTime: 2022-07-21 21:52:31
+Discription: 
+'''
+import sys,os
+curr_path = os.path.dirname(os.path.abspath(__file__))  # current path
+parent_path = os.path.dirname(curr_path)  # parent path
+sys.path.append(parent_path)  # add to system path
+
+import gym
+import torch
+import datetime
+import argparse
+
+from common.utils import save_results,make_dir
+from common.utils import plot_rewards,save_args
+from DoubleDQN.double_dqn import DoubleDQN
+
+def get_args():
+    """ Hyperparameters
+    """
+    curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")  # Obtain current time
+    parser = argparse.ArgumentParser(description="hyperparameters")      
+    parser.add_argument('--algo_name',default='DoubleDQN',type=str,help="name of algorithm")
+    parser.add_argument('--env_name',default='CartPole-v0',type=str,help="name of environment")
+    parser.add_argument('--train_eps',default=200,type=int,help="episodes of training")
+    parser.add_argument('--test_eps',default=20,type=int,help="episodes of testing")
+    parser.add_argument('--gamma',default=0.99,type=float,help="discounted factor")
+    parser.add_argument('--epsilon_start',default=0.95,type=float,help="initial value of epsilon")
+    parser.add_argument('--epsilon_end',default=0.01,type=float,help="final value of epsilon")
+    parser.add_argument('--epsilon_decay',default=500,type=int,help="decay rate of epsilon")
+    parser.add_argument('--lr',default=0.0001,type=float,help="learning rate")
+    parser.add_argument('--memory_capacity',default=100000,type=int,help="memory capacity")
+    parser.add_argument('--batch_size',default=64,type=int)
+    parser.add_argument('--target_update',default=2,type=int)
+    parser.add_argument('--hidden_dim',default=256,type=int)
+    parser.add_argument('--device',default='cpu',type=str,help="cpu or cuda") 
+    parser.add_argument('--result_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/results/' )
+    parser.add_argument('--model_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/models/' ) # path to save models
+    parser.add_argument('--save_fig',default=True,type=bool,help="if save figure or not")           
+    args = parser.parse_args()                          
+    return args
+        
+        
+def env_agent_config(cfg,seed=1):
+    env = gym.make(cfg.env_name)  
+    env.seed(seed)
+    n_states = env.observation_space.shape[0]
+    n_actions = env.action_space.n
+    agent = DoubleDQN(n_states,n_actions,cfg)
+    return env,agent
+
+def train(cfg,env,agent):
+    print('Start training!')
+    print(f'Env:{cfg.env_name}, Algorithm:{cfg.algo_name}, Device:{cfg.device}')
+    rewards = [] # 记录所有回合的奖励
+    ma_rewards = []  # 记录所有回合的滑动平均奖励
+    for i_ep in range(cfg.train_eps):
+        ep_reward = 0 # 记录一回合内的奖励
+        state = env.reset() # 重置环境，返回初始状态
+        while True:
+            action = agent.choose_action(state) 
+            next_state, reward, done, _ = env.step(action)
+            ep_reward += reward
+            agent.memory.push(state, action, reward, next_state, done) 
+            state = next_state 
+            agent.update() 
+            if done:
+                break
+        if i_ep % cfg.target_update == 0:
+            agent.target_net.load_state_dict(agent.policy_net.state_dict())
+        if (i_ep+1)%10 == 0: 
+            print(f'Env:{i_ep+1}/{cfg.train_eps}, Reward:{ep_reward:.2f}')
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(
+                0.9*ma_rewards[-1]+0.1*ep_reward)
+        else:
+            ma_rewards.append(ep_reward)   
+    print('Finish training!')
+    return {'rewards':rewards,'ma_rewards':ma_rewards}
+
+def test(cfg,env,agent):
+    print('Start testing')
+    print(f'Env:{cfg.env_name}, Algorithm:{cfg.algo_name}, Device:{cfg.device}')
+    ############# 由于测试不需要使用epsilon-greedy策略，所以相应的值设置为0 ###############
+    cfg.epsilon_start = 0.0  # e-greedy策略中初始epsilon
+    cfg.epsilon_end = 0.0  # e-greedy策略中的终止epsilon
+    ################################################################################
+    rewards = [] # 记录所有回合的奖励
+    ma_rewards = []  # 记录所有回合的滑动平均奖励
+    
+    for i_ep in range(cfg.test_eps):
+        state = env.reset() 
+        ep_reward = 0   
+        while True:
+            action = agent.choose_action(state) 
+            next_state, reward, done, _ = env.step(action)  
+            state = next_state  
+            ep_reward += reward
+            if done:
+                break
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(ma_rewards[-1]*0.9+ep_reward*0.1)
+        else:
+            ma_rewards.append(ep_reward)
+        print(f"Epside:{i_ep+1}/{cfg.test_eps}, Reward:{ep_reward:.1f}")
+    print('Finish testing!')
+    return {'rewards':rewards,'ma_rewards':ma_rewards}
+
+if __name__ == "__main__":
+    cfg = get_args()
+    print(cfg.device)
+    # training
+    env,agent = env_agent_config(cfg,seed=1)
+    res_dic = train(cfg, env, agent)
+    make_dir(cfg.result_path, cfg.model_path)
+    save_args(cfg)
+    agent.save(path=cfg.model_path)
+    save_results(res_dic, tag='train',
+                 path=cfg.result_path)  
+    plot_rewards(res_dic['rewards'], res_dic['ma_rewards'], cfg, tag="train")  
+    # testing
+    env,agent = env_agent_config(cfg,seed=10)
+    agent.load(path=cfg.model_path)
+    res_dic = test(cfg,env,agent)
+    save_results(res_dic, tag='test',
+                 path=cfg.result_path)  
+    plot_rewards(res_dic['rewards'], res_dic['ma_rewards'], cfg, tag="test") 

projects/codes/GAE/task0_train.py

@@ -0,0 +1,167 @@
+import math
+import random
+
+import gym
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import torch.nn.functional as F
+from torch.distributions import Normal
+import matplotlib.pyplot as plt
+import seaborn as sns
+import sys,os
+curr_path = os.path.dirname(os.path.abspath(__file__)) # 当前文件所在绝对路径
+parent_path = os.path.dirname(curr_path) # 父路径
+sys.path.append(parent_path) # 添加父路径到系统路径sys.path
+
+use_cuda = torch.cuda.is_available()
+device   = torch.device("cuda" if use_cuda else "cpu")
+
+from common.multiprocessing_env import SubprocVecEnv
+
+num_envs = 16
+env_name = "Pendulum-v0"
+
+def make_env():
+    def _thunk():
+        env = gym.make(env_name)
+        return env
+
+    return _thunk
+
+envs = [make_env() for i in range(num_envs)]
+envs = SubprocVecEnv(envs)
+
+env = gym.make(env_name)
+
+def init_weights(m):
+    if isinstance(m, nn.Linear):
+        nn.init.normal_(m.weight, mean=0., std=0.1)
+        nn.init.constant_(m.bias, 0.1)
+
+class ActorCritic(nn.Module):
+    def __init__(self, num_inputs, num_outputs, hidden_size, std=0.0):
+        super(ActorCritic, self).__init__()
+        
+        self.critic = nn.Sequential(
+            nn.Linear(num_inputs, hidden_size),
+            nn.ReLU(),
+            nn.Linear(hidden_size, 1)
+        )
+        
+        self.actor = nn.Sequential(
+            nn.Linear(num_inputs, hidden_size),
+            nn.ReLU(),
+            nn.Linear(hidden_size, num_outputs),
+        )
+        self.log_std = nn.Parameter(torch.ones(1, num_outputs) * std)
+        
+        self.apply(init_weights)
+        
+    def forward(self, x):
+        value = self.critic(x)
+        mu    = self.actor(x)
+        std   = self.log_std.exp().expand_as(mu)
+        dist  = Normal(mu, std)
+        return dist, value
+
+
+def plot(frame_idx, rewards):
+    plt.figure(figsize=(20,5))
+    plt.subplot(131)
+    plt.title('frame %s. reward: %s' % (frame_idx, rewards[-1]))
+    plt.plot(rewards)
+    plt.show()
+    
+def test_env(vis=False):
+    state = env.reset()
+    if vis: env.render()
+    done = False
+    total_reward = 0
+    while not done:
+        state = torch.FloatTensor(state).unsqueeze(0).to(device)
+        dist, _ = model(state)
+        next_state, reward, done, _ = env.step(dist.sample().cpu().numpy()[0])
+        state = next_state
+        if vis: env.render()
+        total_reward += reward
+    return total_reward
+
+def compute_gae(next_value, rewards, masks, values, gamma=0.99, tau=0.95):
+    values = values + [next_value]
+    gae = 0
+    returns = []
+    for step in reversed(range(len(rewards))):
+        delta = rewards[step] + gamma * values[step + 1] * masks[step] - values[step]
+        gae = delta + gamma * tau * masks[step] * gae
+        returns.insert(0, gae + values[step])
+    return returns
+
+num_inputs  = envs.observation_space.shape[0]
+num_outputs = envs.action_space.shape[0]
+
+#Hyper params:
+hidden_size = 256
+lr          = 3e-2
+num_steps   = 20
+
+model = ActorCritic(num_inputs, num_outputs, hidden_size).to(device)
+optimizer = optim.Adam(model.parameters())
+
+max_frames   = 100000
+frame_idx    = 0
+test_rewards = []
+
+state = envs.reset()
+
+while frame_idx < max_frames:
+
+    log_probs = []
+    values    = []
+    rewards   = []
+    masks     = []
+    entropy = 0
+
+    for _ in range(num_steps):
+        state = torch.FloatTensor(state).to(device)
+        dist, value = model(state)
+
+        action = dist.sample()
+        next_state, reward, done, _ = envs.step(action.cpu().numpy())
+
+        log_prob = dist.log_prob(action)
+        entropy += dist.entropy().mean()
+        
+        log_probs.append(log_prob)
+        values.append(value)
+        rewards.append(torch.FloatTensor(reward).unsqueeze(1).to(device))
+        masks.append(torch.FloatTensor(1 - done).unsqueeze(1).to(device))
+        
+        state = next_state
+        frame_idx += 1
+        
+        if frame_idx % 1000 == 0:
+            test_rewards.append(np.mean([test_env() for _ in range(10)]))
+            print(test_rewards[-1])
+            # plot(frame_idx, test_rewards)
+            
+    next_state = torch.FloatTensor(next_state).to(device)
+    _, next_value = model(next_state)
+    returns = compute_gae(next_value, rewards, masks, values)
+    
+    log_probs = torch.cat(log_probs)
+    returns   = torch.cat(returns).detach()
+    values    = torch.cat(values)
+
+    advantage = returns - values
+
+    actor_loss  = -(log_probs * advantage.detach()).mean()
+    critic_loss = advantage.pow(2).mean()
+
+    loss = actor_loss + 0.5 * critic_loss - 0.001 * entropy
+
+    optimizer.zero_grad()
+    loss.backward()
+    optimizer.step()

projects/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)

projects/codes/HierarchicalDQN/agent.py

@@ -0,0 +1,154 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2021-03-24 22:18:18
+LastEditor: John
+LastEditTime: 2021-05-04 22:39:34
+Discription: 
+Environment: 
+'''
+import torch
+import torch.nn as nn
+import torch.optim as optim
+import torch.nn.functional as F
+import numpy as np
+import random,math
+
+class ReplayBuffer:
+    def __init__(self, capacity):
+        self.capacity = capacity # 经验回放的容量
+        self.buffer = [] # 缓冲区
+        self.position = 0 
+    
+    def push(self, state, action, reward, next_state, done):
+        ''' 缓冲区是一个队列，容量超出时去掉开始存入的转移(transition)
+        '''
+        if len(self.buffer) < self.capacity:
+            self.buffer.append(None)
+        self.buffer[self.position] = (state, action, reward, next_state, done)
+        self.position = (self.position + 1) % self.capacity 
+    
+    def sample(self, batch_size):
+        batch = random.sample(self.buffer, batch_size) # 随机采出小批量转移
+        state, action, reward, next_state, done =  zip(*batch) # 解压成状态，动作等
+        return state, action, reward, next_state, done
+    
+    def __len__(self):
+        ''' 返回当前存储的量
+        '''
+        return len(self.buffer)
+class MLP(nn.Module):
+    def __init__(self, input_dim,output_dim,hidden_dim=128):
+        """ 初始化q网络，为全连接网络
+            input_dim: 输入的特征数即环境的状态维度
+            output_dim: 输出的动作维度
+        """
+        super(MLP, self).__init__()
+        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, output_dim) # 输出层
+        
+    def forward(self, x):
+        # 各层对应的激活函数
+        x = F.relu(self.fc1(x)) 
+        x = F.relu(self.fc2(x))
+        return self.fc3(x)
+        
+class HierarchicalDQN:
+    def __init__(self,n_states,n_actions,cfg):
+        self.n_states = n_states
+        self.n_actions = n_actions
+        self.gamma = cfg.gamma
+        self.device = cfg.device
+        self.batch_size = cfg.batch_size
+        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.policy_net = MLP(2*n_states, n_actions,cfg.hidden_dim).to(self.device)
+        self.meta_policy_net = MLP(n_states, n_states,cfg.hidden_dim).to(self.device)
+        self.optimizer = optim.Adam(self.policy_net.parameters(),lr=cfg.lr)
+        self.meta_optimizer = optim.Adam(self.meta_policy_net.parameters(),lr=cfg.lr)
+        self.memory = ReplayBuffer(cfg.memory_capacity)
+        self.meta_memory = ReplayBuffer(cfg.memory_capacity)
+        self.loss_numpy  = 0
+        self.meta_loss_numpy  = 0
+        self.losses = []
+        self.meta_losses = []
+    def to_onehot(self,x):
+        oh = np.zeros(self.n_states)
+        oh[x - 1] = 1.
+        return oh
+    def set_goal(self,state):
+        if random.random() > self.epsilon(self.frame_idx):
+            with torch.no_grad():
+                state = torch.tensor(state, device=self.device, dtype=torch.float32).unsqueeze(0)
+                goal = self.meta_policy_net(state).max(1)[1].item() 
+        else:
+            goal = random.randrange(self.n_states)
+        return goal
+    def choose_action(self,state):
+        self.frame_idx += 1
+        if random.random() > self.epsilon(self.frame_idx):
+            with torch.no_grad():
+                state = torch.tensor(state, device=self.device, dtype=torch.float32).unsqueeze(0)
+                q_value = self.policy_net(state)
+                action = q_value.max(1)[1].item()  
+        else:
+            action = random.randrange(self.n_actions)
+        return action
+    def update(self):
+        self.update_policy()
+        self.update_meta()
+    def update_policy(self): 
+        if self.batch_size > len(self.memory):
+            return
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.memory.sample(self.batch_size)
+        state_batch = torch.tensor(state_batch,device=self.device,dtype=torch.float)
+        action_batch = torch.tensor(action_batch,device=self.device,dtype=torch.int64).unsqueeze(1)  
+        reward_batch = torch.tensor(reward_batch,device=self.device,dtype=torch.float)  
+        next_state_batch = torch.tensor(next_state_batch,device=self.device, dtype=torch.float)
+        done_batch = torch.tensor(np.float32(done_batch),device=self.device)
+        q_values = self.policy_net(state_batch).gather(dim=1, index=action_batch).squeeze(1)
+        next_state_values = self.policy_net(next_state_batch).max(1)[0].detach()
+        expected_q_values = reward_batch + 0.99 * next_state_values * (1-done_batch)
+        loss = nn.MSELoss()(q_values, expected_q_values) 
+        self.optimizer.zero_grad() 
+        loss.backward()
+        for param in self.policy_net.parameters():  # clip防止梯度爆炸
+            param.grad.data.clamp_(-1, 1)
+        self.optimizer.step()  
+        self.loss_numpy = loss.detach().cpu().numpy()
+        self.losses.append(self.loss_numpy)  
+    def update_meta(self):
+        if self.batch_size > len(self.meta_memory):
+            return
+        state_batch, action_batch, reward_batch, next_state_batch, done_batch = self.meta_memory.sample(self.batch_size)
+        state_batch = torch.tensor(state_batch,device=self.device,dtype=torch.float)
+        action_batch = torch.tensor(action_batch,device=self.device,dtype=torch.int64).unsqueeze(1)  
+        reward_batch = torch.tensor(reward_batch,device=self.device,dtype=torch.float)  
+        next_state_batch = torch.tensor(next_state_batch,device=self.device, dtype=torch.float)
+        done_batch = torch.tensor(np.float32(done_batch),device=self.device)
+        q_values = self.meta_policy_net(state_batch).gather(dim=1, index=action_batch).squeeze(1)
+        next_state_values = self.meta_policy_net(next_state_batch).max(1)[0].detach()
+        expected_q_values = reward_batch + 0.99 * next_state_values * (1-done_batch)
+        meta_loss = nn.MSELoss()(q_values, expected_q_values) 
+        self.meta_optimizer.zero_grad() 
+        meta_loss.backward()
+        for param in self.meta_policy_net.parameters():  # clip防止梯度爆炸
+            param.grad.data.clamp_(-1, 1)
+        self.meta_optimizer.step() 
+        self.meta_loss_numpy = meta_loss.detach().cpu().numpy()
+        self.meta_losses.append(self.meta_loss_numpy)
+
+    def save(self, path):
+        torch.save(self.policy_net.state_dict(), path+'policy_checkpoint.pth')
+        torch.save(self.meta_policy_net.state_dict(), path+'meta_checkpoint.pth')
+
+    def load(self, path):
+        self.policy_net.load_state_dict(torch.load(path+'policy_checkpoint.pth'))
+        self.meta_policy_net.load_state_dict(torch.load(path+'meta_checkpoint.pth'))
+        
+
+        
+        

projects/codes/HierarchicalDQN/task0.py

@@ -0,0 +1,88 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2021-03-29 10:37:32
+LastEditor: John
+LastEditTime: 2021-05-04 22:35:56
+Discription: 
+Environment: 
+'''
+import sys
+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 datetime
+import numpy as np
+import torch
+import gym
+
+from common.utils import save_results,make_dir
+from common.utils import plot_rewards
+from HierarchicalDQN.agent import HierarchicalDQN
+from HierarchicalDQN.train import train,test
+
+curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")  # 获取当前时间
+algo_name = "Hierarchical DQN"  # 算法名称
+env_name = 'CartPole-v0'  # 环境名称
+class HierarchicalDQNConfig:
+    def __init__(self):
+        self.algo_name = algo_name  # 算法名称
+        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episode数目
+        self.test_eps = 50  # 测试的episode数目
+        self.gamma = 0.99
+        self.epsilon_start = 1  # start epsilon of e-greedy policy
+        self.epsilon_end = 0.01
+        self.epsilon_decay = 200
+        self.lr = 0.0001  # learning rate
+        self.memory_capacity = 10000  # Replay Memory capacity
+        self.batch_size = 32
+        self.target_update = 2  # 目标网络的更新频率
+        self.hidden_dim = 256  # 网络隐藏层
+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 = torch.device(
+            "cuda" if torch.cuda.is_available() else "cpu")  # 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):
+    env = gym.make(cfg.env_name)  
+    env.seed(seed)
+    n_states = env.observation_space.shape[0]
+    n_actions = env.action_space.n
+    agent = HierarchicalDQN(n_states,n_actions,cfg)
+    return env,agent
+
+if __name__ == "__main__":
+    cfg = HierarchicalDQNConfig()
+    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
+

projects/codes/HierarchicalDQN/train.py

@@ -0,0 +1,77 @@
+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 numpy as np
+
+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 = [] # 记录所有回合的奖励
+    ma_rewards = []  # 记录所有回合的滑动平均奖励
+    for i_ep in range(cfg.train_eps):
+        state = env.reset()
+        done = False
+        ep_reward = 0
+        while not done:
+            goal = agent.set_goal(state)
+            onehot_goal = agent.to_onehot(goal)
+            meta_state = state
+            extrinsic_reward = 0
+            while not done and goal != np.argmax(state):
+                goal_state = np.concatenate([state, onehot_goal])
+                action = agent.choose_action(goal_state)
+                next_state, reward, done, _ = env.step(action)
+                ep_reward += reward
+                extrinsic_reward += reward
+                intrinsic_reward = 1.0 if goal == np.argmax(
+                    next_state) else 0.0
+                agent.memory.push(goal_state, action, intrinsic_reward, np.concatenate(
+                    [next_state, onehot_goal]), done)
+                state = next_state
+                agent.update()
+        if (i_ep+1)%10 == 0: 
+            print(f'回合：{i_ep+1}/{cfg.train_eps}，奖励：{ep_reward}，Loss:{agent.loss_numpy:.2f}， Meta_Loss:{agent.meta_loss_numpy:.2f}')
+        agent.meta_memory.push(meta_state, goal, extrinsic_reward, state, done)
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(
+                0.9*ma_rewards[-1]+0.1*ep_reward)
+        else:
+            ma_rewards.append(ep_reward)
+    print('完成训练！')
+    return rewards, ma_rewards
+
+def 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}')
+    rewards = [] # 记录所有回合的奖励
+    ma_rewards = []  # 记录所有回合的滑动平均奖励
+    for i_ep in range(cfg.train_eps):
+        state = env.reset()
+        done = False
+        ep_reward = 0
+        while not done:
+            goal = agent.set_goal(state)
+            onehot_goal = agent.to_onehot(goal)
+            extrinsic_reward = 0
+            while not done and goal != np.argmax(state):
+                goal_state = np.concatenate([state, onehot_goal])
+                action = agent.choose_action(goal_state)
+                next_state, reward, done, _ = env.step(action)
+                ep_reward += reward
+                extrinsic_reward += reward
+                state = next_state
+                agent.update()
+        if (i_ep+1)%10 == 0: 
+            print(f'回合：{i_ep+1}/{cfg.train_eps}，奖励：{ep_reward}，Loss:{agent.loss_numpy:.2f}， Meta_Loss:{agent.meta_loss_numpy:.2f}')
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(
+                0.9*ma_rewards[-1]+0.1*ep_reward)
+        else:
+            ma_rewards.append(ep_reward)
+    print('完成训练！')
+    return rewards, ma_rewards

projects/codes/MonteCarlo/README.md

@@ -0,0 +1,5 @@
+# *On-Policy First-Visit MC Control*
+
+### 伪代码
+
+![mc_control_algo](assets/mc_control_algo.png)

projects/codes/MonteCarlo/agent.py

@@ -0,0 +1,65 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2021-03-12 16:14:34
+LastEditor: John
+LastEditTime: 2021-05-05 16:58:39
+Discription: 
+Environment: 
+'''
+import numpy as np
+from collections import defaultdict
+import torch
+import dill
+
+class FisrtVisitMC:
+    ''' On-Policy First-Visit MC Control
+    '''
+    def __init__(self,n_actions,cfg):
+        self.n_actions = n_actions
+        self.epsilon = cfg.epsilon
+        self.gamma = cfg.gamma 
+        self.Q_table = defaultdict(lambda: np.zeros(n_actions))
+        self.returns_sum = defaultdict(float) # sum of returns
+        self.returns_count = defaultdict(float)
+        
+    def choose_action(self,state):
+        ''' e-greed policy '''
+        if state in self.Q_table.keys():
+            best_action = np.argmax(self.Q_table[state])
+            action_probs = np.ones(self.n_actions, dtype=float) * self.epsilon / self.n_actions
+            action_probs[best_action] += (1.0 - self.epsilon)
+            action = np.random.choice(np.arange(len(action_probs)), p=action_probs)
+        else:
+            action = np.random.randint(0,self.n_actions)
+        return action
+    def update(self,one_ep_transition):
+        # Find all (state, action) pairs we've visited in this one_ep_transition
+        # We convert each state to a tuple so that we can use it as a dict key
+        sa_in_episode = set([(tuple(x[0]), x[1]) for x in one_ep_transition])
+        for state, action in sa_in_episode:
+            sa_pair = (state, action)
+            # Find the first occurence of the (state, action) pair in the one_ep_transition
+            first_occurence_idx = next(i for i,x in enumerate(one_ep_transition)
+                                       if x[0] == state and x[1] == action)
+            # Sum up all rewards since the first occurance
+            G = sum([x[2]*(self.gamma**i) for i,x in enumerate(one_ep_transition[first_occurence_idx:])])
+            # Calculate average return for this state over all sampled episodes
+            self.returns_sum[sa_pair] += G
+            self.returns_count[sa_pair] += 1.0
+            self.Q_table[state][action] = self.returns_sum[sa_pair] / self.returns_count[sa_pair]
+    def save(self,path):
+        '''把 Q表格 的数据保存到文件中
+        '''
+        torch.save(
+            obj=self.Q_table,
+            f=path+"Q_table",
+            pickle_module=dill
+        )
+
+    def load(self, path):
+        '''从文件中读取数据到 Q表格
+        '''
+        self.Q_table =torch.load(f=path+"Q_table",pickle_module=dill)

projects/codes/MonteCarlo/task0_train.py

@@ -0,0 +1,118 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2021-03-11 14:26:44
+LastEditor: John
+LastEditTime: 2021-05-05 17:27:50
+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 torch
+import datetime
+
+from common.utils import save_results,make_dir
+from common.plot import plot_rewards
+from MonteCarlo.agent import FisrtVisitMC
+from envs.racetrack_env import RacetrackEnv
+
+curr_time = datetime.datetime.now().strftime(
+    "%Y%m%d-%H%M%S")  # obtain current time
+
+class MCConfig:
+    def __init__(self):
+        self.algo = "MC"  # name of algo
+        self.env = 'Racetrack'
+        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
+        # epsilon: The probability to select a random action .
+        self.epsilon = 0.15
+        self.gamma = 0.9  # gamma: Gamma discount factor.
+        self.train_eps = 200
+        self.device = torch.device(
+            "cuda" if torch.cuda.is_available() else "cpu")  # check gpu
+
+def env_agent_config(cfg,seed=1):
+    env = RacetrackEnv()
+    n_actions = 9
+    agent = FisrtVisitMC(n_actions, cfg)
+    return env,agent
+    
+def train(cfg, env, agent):
+    print('Start to eval !')
+    print(f'Env:{cfg.env}, Algorithm:{cfg.algo}, Device:{cfg.device}')
+    rewards = []
+    ma_rewards = []  # moving average rewards
+    for i_ep in range(cfg.train_eps):
+        state = env.reset()
+        ep_reward = 0
+        one_ep_transition = []
+        while True:
+            action = agent.choose_action(state)
+            next_state, reward, done = env.step(action)
+            ep_reward += reward
+            one_ep_transition.append((state, action, reward))
+            state = next_state
+            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)
+        agent.update(one_ep_transition)
+        if (i_ep+1) % 10 == 0:
+            print(f"Episode:{i_ep+1}/{cfg.train_eps}: Reward:{ep_reward}")
+    print('Complete training！')
+    return rewards, ma_rewards
+
+def eval(cfg, env, agent):
+    print('Start to eval !')
+    print(f'Env:{cfg.env}, Algorithm:{cfg.algo}, Device:{cfg.device}')
+    rewards = []
+    ma_rewards = []  # moving average rewards
+    for i_ep in range(cfg.train_eps):
+        state = env.reset()
+        ep_reward = 0
+        while True:
+            action = agent.choose_action(state)
+            next_state, reward, done = env.step(action)
+            ep_reward += reward
+            state = next_state
+            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 (i_ep+1) % 10 == 0:
+            print(f"Episode:{i_ep+1}/{cfg.train_eps}: Reward:{ep_reward}")
+    return rewards, ma_rewards
+    
+if __name__ == "__main__":
+    cfg = MCConfig()
+    
+    # train
+    env,agent = env_agent_config(cfg,seed=1)
+    rewards, ma_rewards = train(cfg, env, agent)
+    make_dir(cfg.result_path, cfg.model_path)
+    agent.save(path=cfg.model_path)
+    save_results(rewards, ma_rewards, tag='train', path=cfg.result_path)
+    plot_rewards(rewards, ma_rewards, tag="train",
+                 algo=cfg.algo, path=cfg.result_path)
+    # eval
+    env,agent = env_agent_config(cfg,seed=10)
+    agent.load(path=cfg.model_path)
+    rewards,ma_rewards = eval(cfg,env,agent)
+    save_results(rewards,ma_rewards,tag='eval',path=cfg.result_path)
+    plot_rewards(rewards,ma_rewards,tag="eval",env=cfg.env,algo = cfg.algo,path=cfg.result_path)

projects/codes/NoisyDQN/noisy_dqn.py

@@ -0,0 +1,52 @@
+import torch
+import torch.nn as nn
+
+class NoisyLinear(nn.Module):
+    def __init__(self, input_dim, output_dim, std_init=0.4):
+        super(NoisyLinear, self).__init__()
+        
+        self.input_dim  = input_dim
+        self.output_dim = output_dim
+        self.std_init     = std_init
+        
+        self.weight_mu    = nn.Parameter(torch.FloatTensor(output_dim, input_dim))
+        self.weight_sigma = nn.Parameter(torch.FloatTensor(output_dim, input_dim))
+        self.register_buffer('weight_epsilon', torch.FloatTensor(output_dim, input_dim))
+        
+        self.bias_mu    = nn.Parameter(torch.FloatTensor(output_dim))
+        self.bias_sigma = nn.Parameter(torch.FloatTensor(output_dim))
+        self.register_buffer('bias_epsilon', torch.FloatTensor(output_dim))
+        
+        self.reset_parameters()
+        self.reset_noise()
+    
+    def forward(self, x):
+        if self.training: 
+            weight = self.weight_mu + self.weight_sigma.mul( (self.weight_epsilon))
+            bias   = self.bias_mu   + self.bias_sigma.mul(Variable(self.bias_epsilon))
+        else:
+            weight = self.weight_mu
+            bias   = self.bias_mu
+        
+        return F.linear(x, weight, bias)
+    
+    def reset_parameters(self):
+        mu_range = 1 / math.sqrt(self.weight_mu.size(1))
+        
+        self.weight_mu.data.uniform_(-mu_range, mu_range)
+        self.weight_sigma.data.fill_(self.std_init / math.sqrt(self.weight_sigma.size(1)))
+        
+        self.bias_mu.data.uniform_(-mu_range, mu_range)
+        self.bias_sigma.data.fill_(self.std_init / math.sqrt(self.bias_sigma.size(0)))
+    
+    def reset_noise(self):
+        epsilon_in  = self._scale_noise(self.input_dim)
+        epsilon_out = self._scale_noise(self.output_dim)
+        
+        self.weight_epsilon.copy_(epsilon_out.ger(epsilon_in))
+        self.bias_epsilon.copy_(self._scale_noise(self.output_dim))
+    
+    def _scale_noise(self, size):
+        x = torch.randn(size)
+        x = x.sign().mul(x.abs().sqrt())
+        return x

projects/codes/NoisyDQN/task0_train.ipynb

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

projects/codes/PPO/README.md

@@ -0,0 +1,142 @@
+## 原理简介
+
+PPO是一种on-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,n_states, n_actions,
+            hidden_dim=256):
+        super(Actor, self).__init__()
+
+        self.actor = nn.Sequential(
+                nn.Linear(n_states, hidden_dim),
+                nn.ReLU(),
+                nn.Linear(hidden_dim, hidden_dim),
+                nn.ReLU(),
+                nn.Linear(hidden_dim, n_actions),
+                nn.Softmax(dim=-1)
+        )
+    def forward(self, state):
+        dist = self.actor(state)
+        dist = Categorical(dist)
+        return dist
+
+class Critic(nn.Module):
+    def __init__(self, n_states,hidden_dim=256):
+        super(Critic, self).__init__()
+        self.critic = nn.Sequential(
+                nn.Linear(n_states, 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```n_states+n_actions```，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)

projects/codes/PPO/outputs/CartPole-v0/20220731-233512/results/params.json

@@ -0,0 +1,20 @@
+{
+    "algo_name": "PPO",
+    "env_name": "CartPole-v0",
+    "continuous": false,
+    "train_eps": 200,
+    "test_eps": 20,
+    "gamma": 0.99,
+    "batch_size": 5,
+    "n_epochs": 4,
+    "actor_lr": 0.0003,
+    "critic_lr": 0.0003,
+    "gae_lambda": 0.95,
+    "policy_clip": 0.2,
+    "update_fre": 20,
+    "hidden_dim": 256,
+    "device": "cpu",
+    "result_path": "C:\\Users\\24438\\Desktop\\rl-tutorials\\codes\\PPO/outputs/CartPole-v0/20220731-233512/results/",
+    "model_path": "C:\\Users\\24438\\Desktop\\rl-tutorials\\codes\\PPO/outputs/CartPole-v0/20220731-233512/models/",
+    "save_fig": true
+}

projects/codes/PPO/ppo2.py

@@ -0,0 +1,162 @@
+#!/usr/bin/env python
+# coding=utf-8
+'''
+Author: John
+Email: johnjim0816@gmail.com
+Date: 2021-03-23 15:17:42
+LastEditor: John
+LastEditTime: 2021-12-31 19:38:33
+Discription: 
+Environment: 
+'''
+import os
+import numpy as np
+import torch 
+import torch.optim as optim
+import torch.nn as nn
+from torch.distributions.categorical import Categorical
+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 = []
+class Actor(nn.Module):
+    def __init__(self,n_states, n_actions,
+            hidden_dim):
+        super(Actor, self).__init__()
+
+        self.actor = nn.Sequential(
+                nn.Linear(n_states, hidden_dim),
+                nn.ReLU(),
+                nn.Linear(hidden_dim, hidden_dim),
+                nn.ReLU(),
+                nn.Linear(hidden_dim, n_actions),
+                nn.Softmax(dim=-1)
+        )
+    def forward(self, state):
+        dist = self.actor(state)
+        dist = Categorical(dist)
+        return dist
+
+class Critic(nn.Module):
+    def __init__(self, n_states,hidden_dim):
+        super(Critic, self).__init__()
+        self.critic = nn.Sequential(
+                nn.Linear(n_states, 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
+class PPO:
+    def __init__(self, n_states, n_actions,cfg):
+        self.gamma = cfg.gamma
+        self.continuous = cfg.continuous 
+        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(n_states, n_actions,cfg.hidden_dim).to(self.device)
+        self.critic = Critic(n_states,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, state):
+        state = np.array([state]) # 先转成数组再转tensor更高效
+        state = torch.tensor(state, dtype=torch.float).to(self.device)
+        dist = self.actor(state)
+        value = self.critic(state)
+        action = dist.sample()
+        probs = torch.squeeze(dist.log_prob(action)).item()
+        if self.continuous:
+            action = torch.tanh(action)
+        else:
+            action = torch.squeeze(action).item()
+        value = torch.squeeze(value).item()
+        return action, probs, value
+
+    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.loss  = total_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()  
+    def save(self,path):
+        actor_checkpoint = os.path.join(path, 'ppo_actor.pt')
+        critic_checkpoint= os.path.join(path, 'ppo_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, 'ppo_actor.pt')
+        critic_checkpoint= os.path.join(path, 'ppo_critic.pt')
+        self.actor.load_state_dict(torch.load(actor_checkpoint)) 
+        self.critic.load_state_dict(torch.load(critic_checkpoint))  
+
+

projects/codes/PPO/task0.py

@@ -0,0 +1,132 @@
+import sys,os
+curr_path = os.path.dirname(os.path.abspath(__file__)) # 当前文件所在绝对路径
+parent_path = os.path.dirname(curr_path) # 父路径
+sys.path.append(parent_path) # 添加路径到系统路径
+
+import gym
+import torch
+import numpy as np
+import datetime
+import argparse
+from common.utils import plot_rewards,save_args,save_results,make_dir
+from ppo2 import PPO
+
+def get_args():
+    """ Hyperparameters
+    """
+    curr_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")  # 获取当前时间
+    parser = argparse.ArgumentParser(description="hyperparameters")      
+    parser.add_argument('--algo_name',default='PPO',type=str,help="name of algorithm")
+    parser.add_argument('--env_name',default='CartPole-v0',type=str,help="name of environment")
+    parser.add_argument('--continuous',default=False,type=bool,help="if PPO is continous") # PPO既可适用于连续动作空间，也可以适用于离散动作空间
+    parser.add_argument('--train_eps',default=200,type=int,help="episodes of training")
+    parser.add_argument('--test_eps',default=20,type=int,help="episodes of testing")
+    parser.add_argument('--gamma',default=0.99,type=float,help="discounted factor")
+    parser.add_argument('--batch_size',default=5,type=int) # mini-batch SGD中的批量大小
+    parser.add_argument('--n_epochs',default=4,type=int)
+    parser.add_argument('--actor_lr',default=0.0003,type=float,help="learning rate of actor net")
+    parser.add_argument('--critic_lr',default=0.0003,type=float,help="learning rate of critic net")
+    parser.add_argument('--gae_lambda',default=0.95,type=float)
+    parser.add_argument('--policy_clip',default=0.2,type=float) # PPO-clip中的clip参数，一般是0.1~0.2左右
+    parser.add_argument('--update_fre',default=20,type=int)
+    parser.add_argument('--hidden_dim',default=256,type=int)
+    parser.add_argument('--device',default='cpu',type=str,help="cpu or cuda") 
+    parser.add_argument('--result_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/results/' )
+    parser.add_argument('--model_path',default=curr_path + "/outputs/" + parser.parse_args().env_name + \
+            '/' + curr_time + '/models/' ) # path to save models
+    parser.add_argument('--save_fig',default=True,type=bool,help="if save figure or not")           
+    args = parser.parse_args()                          
+    return args
+        
+def env_agent_config(cfg,seed = 1):
+    ''' 创建环境和智能体
+    '''
+    env = gym.make(cfg.env_name)  # 创建环境
+    n_states = env.observation_space.shape[0]  # 状态维度
+    if cfg.continuous:
+        n_actions = env.action_space.shape[0] # 动作维度
+    else:
+        n_actions = env.action_space.n  # 动作维度
+    agent = PPO(n_states, n_actions, cfg)  # 创建智能体
+    if seed !=0: # 设置随机种子
+        torch.manual_seed(seed)
+        env.seed(seed)
+        np.random.seed(seed)
+    return env, agent
+
+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 = [] # 记录所有回合的奖励
+    ma_rewards = []  # 记录所有回合的滑动平均奖励
+    steps = 0
+    for i_ep in range(cfg.train_eps):
+        state = env.reset()
+        done = False
+        ep_reward = 0
+        while not done:
+            action, prob, val = agent.choose_action(state)
+            state_, reward, done, _ = env.step(action)
+            steps += 1
+            ep_reward += reward
+            agent.memory.push(state, action, prob, val, reward, done)
+            if steps % cfg.update_fre == 0:
+                agent.update()
+            state = state_
+        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(f"回合：{i_ep+1}/{cfg.train_eps}，奖励：{ep_reward:.2f}")
+    print('完成训练！')
+    env.close()
+    res_dic = {'rewards':rewards,'ma_rewards':ma_rewards}
+    return res_dic
+
+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}')
+    rewards = [] # 记录所有回合的奖励
+    ma_rewards = []  # 记录所有回合的滑动平均奖励
+    for i_ep in range(cfg.test_eps):
+        state = env.reset()
+        done = False
+        ep_reward = 0
+        while not done:
+            action, prob, val = agent.choose_action(state)
+            state_, reward, done, _ = env.step(action)
+            ep_reward += reward
+            state = state_
+        rewards.append(ep_reward)
+        if ma_rewards:
+            ma_rewards.append(
+                0.9*ma_rewards[-1]+0.1*ep_reward)
+        else:
+            ma_rewards.append(ep_reward)
+        print('回合：{}/{}, 奖励：{}'.format(i_ep+1, cfg.test_eps, ep_reward))
+    print('完成训练！')
+    env.close()
+    res_dic = {'rewards':rewards,'ma_rewards':ma_rewards}
+    return res_dic
+
+if __name__ == "__main__":
+    cfg = get_args()
+    # 训练
+    env, agent = env_agent_config(cfg)
+    res_dic = train(cfg, env, agent)
+    make_dir(cfg.result_path, cfg.model_path)  
+    save_args(cfg) # 保存参数
+    agent.save(path=cfg.model_path)  # save model
+    save_results(res_dic, tag='train',
+                 path=cfg.result_path)  
+    plot_rewards(res_dic['rewards'], res_dic['ma_rewards'], cfg, tag="train")  
+    # 测试
+    env, agent = env_agent_config(cfg)
+    agent.load(path=cfg.model_path)  # 导入模型
+    res_dic = test(cfg, env, agent)
+    save_results(res_dic, tag='test',
+                 path=cfg.result_path)  # 保存结果
+    plot_rewards(res_dic['rewards'], res_dic['ma_rewards'],cfg, tag="test")  # 画出结果