optimize reinforce_invpend_gym_v26.py (#638)

docs/tutorials/training_agents/reinforce_invpend_gym_v26.py

@@ -193,29 +193,34 @@ class REINFORCE:
         """Updates the policy network's weights."""
         running_g = 0
         gs = []
-
+    
         # Discounted return (backwards) - [::-1] will return an array in reverse
         for R in self.rewards[::-1]:
             running_g = R + self.gamma * running_g
             gs.insert(0, running_g)
-
+    
         deltas = torch.tensor(gs)
-
-        loss = 0
-        # minimize -1 * prob * reward obtained
-        for log_prob, delta in zip(self.probs, deltas):
-            loss += log_prob.mean() * delta * (-1)
-
+    
+        log_probs = torch.stack(self.probs)
+    
+        # Calculate the mean of log probabilities for all actions in the episode
+        log_prob_mean = log_probs.mean()
+    
+        # Update the loss with the mean log probability and deltas
+        # Now, we compute the correct total loss by taking the sum of the element-wise products.
+        loss = -torch.sum(log_prob_mean * deltas)
+    
         # Update the policy network
         self.optimizer.zero_grad()
         loss.backward()
         self.optimizer.step()
-
+    
         # Empty / zero out all episode-centric/related variables
         self.probs = []
         self.rewards = []
 
 
+
 # %%
 # Now lets train the policy using REINFORCE to master the task of Inverted Pendulum.
 #