用强化学习通关超级马里奥！

def train(cfg, env, agent):
    ''' 训练
    '''
    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):
        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)  # 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
        rewards.append(ep_reward)
        if ma_rewards:
            ma_rewards.append(0.9 * ma_rewards[-1] + 0.1 * ep_reward)
        else:
            ma_rewards.append(ep_reward)
        if (i_ep + 1) % 10 == 0:
            print('回合：{}/{}, 奖励：{}'.format(i_ep + 1, cfg.train_eps, ep_reward))
    print('完成训练！')
    env.close()
    return rewards, ma_rewards

class DQN:
    def __init__(self, state_dim, action_dim, cfg):
        self.action_dim = action_dim  # 总的动作个数
        self.device = cfg.device  # 设备，cpu或gpu等
        self.gamma = cfg.gamma  # 奖励的折扣因子
        # 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(state_dim, action_dim,hidden_dim=cfg.hidden_dim).to(self.device)
        self.target_net = MLP(state_dim, action_dim,hidden_dim=cfg.hidden_dim).to(self.device)
        # 复制参数到目标网路targe_net
        for target_param, param in zip(self.target_net.parameters(),self.policy_net.parameters()): 
            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.action_dim)
        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)
        # 计算当前状态(s,a)对应的Q(s, a)
        q_values = self.policy_net(state_batch).gather(dim=1, index=action_batch) 
        # 计算下一时刻的状态(s_t,a)对应的Q值
        next_q_values = self.target_net(next_state_batch).max(1)[0].detach() 
        # 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()

        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())

        # 计算当前状态(s,a)对应的Q(s, a)
        q_values = self.policy_net(state_batch).gather(dim=1, index=action_batch) 
        # next_q_values = self.target_net(next_state_batch).max(1)[0].detach() 
        # 用policy_net计算下一个状态s_t的最优动作a*
        next_action_batch = torch.argmax(self.policy_net(state_batch),axis=1).unsueeze(1)
        # 用target_net计算下一时刻的状态(s_t_,a)对应的Q值
        next_q_values = self.target_net(next_state_batch).gather(dim=1,index=next_action_batch)
        # 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()

class DuelingNet(nn.Module):
    def __init__(self, state_dim, action_dim,hidden_size=128):
        super(DuelingNet, self).__init__()
        
        # 隐藏层
        self.hidden = nn.Sequential(
            nn.Linear(state_dim, hidden_size),
            nn.ReLU()
        )
        
        # 优势函数
        self.advantage = nn.Sequential(
            nn.Linear(hidden_size, hidden_size),
            nn.ReLU(),
            nn.Linear(hidden_size, action_dim)
        )
        
        # 价值函数
        self.value = nn.Sequential(
            nn.Linear(hidden_size, hidden_size),
            nn.ReLU(),
            nn.Linear(hidden_size, 1)
        )
        
    def forward(self, x):
        x = self.hidden(x)
        advantage = self.advantage(x)
        value     = self.value(x)
 # Q* = A* + V* - mean(A*) 
        return value + advantage  - advantage.mean()