修补bug

eight.py

@@ -70,10 +70,12 @@ def update(self, transition_dict):
 
         q_values = self.q_net(states).gather(1, actions)  # Q值
         # 下个状态的最大Q值
-        max_next_q_values = self.target_q_net(next_states).max(1)[0].view(
-            -1, 1)
-        q_targets = rewards + self.gamma * max_next_q_values * (1 - dones
-                                                                )  # TD误差目标
+        if self.dqn_type == 'DoubleDQN':  # DQN与Double DQN的区别
+            max_action = self.q_net(next_states).max(1)[1].view(-1, 1)
+            max_next_q_values = self.target_q_net(next_states).gather(1, max_action)
+        else:  # DQN的情况
+            max_next_q_values = self.target_q_net(next_states).max(1)[0].view(-1, 1)
+        q_targets = rewards + self.gamma * max_next_q_values * (1 - dones)  # TD误差目标
         dqn_loss = torch.mean(F.mse_loss(q_values, q_targets))  # 均方误差损失函数
         self.optimizer.zero_grad()  # PyTorch中默认梯度会累积,这里需要显式将梯度置为0
         dqn_loss.backward()  # 反向传播更新参数
@@ -84,6 +86,8 @@ def update(self, transition_dict):
                 self.q_net.state_dict())  # 更新目标网络
         self.count += 1
 
+    lr = 1e-2
+
 lr = 1e-2
 num_episodes = 200
 hidden_dim = 128