finally fucking did pacman

.gitignore

@@ -3,3 +3,5 @@
 
 __pycache__/
 *.pyc
+
+q_table.json

04_pacman_rl/pacman.py

@@ -213,7 +213,7 @@ def train(q, num_iterations=10000):
                 elif ghost_y > pacman_y and labyrinth[ghost_y - 1][ghost_x] != "#":
                     ghost_y -= 1
                 
-                s = (pacman_x, pacman_y, ghost_x, ghost_y)
+            s = (pacman_x, pacman_y, ghost_x, ghost_y)
         
         outer_iter += 1
         if outer_iter % 100 == 0:
@@ -276,7 +276,8 @@ def visualize(q, num_games=10):
 
             if iter % 3 == 0:
                 ghost.move_towards_pacman(pacman)
-                s = (pacman.x, pacman.y, ghost.x, ghost.y)
+            
+            s = (pacman.x, pacman.y, ghost.x, ghost.y)
 
             # Draw
             draw_labyrinth()
@@ -284,7 +285,7 @@ def visualize(q, num_games=10):
             ghost.draw()
             pygame.display.flip()
             
-            tick_speed = 20 # if game_num % 20 == 0 else 100
+            tick_speed = 10 # if game_num % 20 == 0 else 100
             clock.tick(tick_speed)
 
 # Main function
@@ -297,7 +298,7 @@ def main():
         q = rl.q_init()
     
     print("Training for 10000 iterations...")
-    q = train(q, num_iterations=20000)
+    q = train(q, num_iterations=10000)
     
     print("\nTraining complete! Starting visualization...")
     visualize(q, num_games=10)

04_pacman_rl/reinforcement_learning.py

@@ -15,7 +15,6 @@ def q_init():
     """ Fill every possible action in every state with a small value for initialization"""
 
     # Configuration
-    NUM_ACTIONS = 4
     RAND_Q_VALUES = [random.uniform(-0.1, 0.1), random.uniform(-0.1, 0.1), random.uniform(-0.1, 0.1), random.uniform(-0.1, 0.1)]
     # print(RAND_Q_VALUES) # debugging
 
@@ -79,35 +78,35 @@ def epsilon_greedy(q, s, epsilon=0.1):
     With probability epsilon, choose a random action. Otherwise choose the greedy action.
     Avoids actions that would result in collision with ghost.
     """
-    if np.random.random() < epsilon:
+    # if np.random.random() < epsilon:
-    # Explore: choose random action (excluding blocked actions with Q=0)
+    # # Explore: choose random action (excluding blocked actions with Q=0)
-        valid_actions = [i for i in range(len(q[s])) if q[s][i] is not None]
+    #     valid_actions = [i for i in range(len(q[s])) if q[s][i] is not None]
-        return np.random.choice(valid_actions)
+    #     return np.random.choice(valid_actions)
         
-    else:
+    # else:
-        # Get all valid (non-blocked) actions with their Q-values
+    # Get all valid (non-blocked) actions with their Q-values
-        valid_actions = [(i, q[s][i]) for i in range(len(q[s])) if q[s][i] is not None]
+    valid_actions = [(i, q[s][i]) for i in range(len(q[s])) if q[s][i] is not None]
     
-        # Sort by Q-value in descending order
+    # Sort by Q-value in descending order
-        valid_actions.sort(key=lambda x: x[1], reverse=True)
+    valid_actions.sort(key=lambda x: x[1], reverse=True)
     
-        # Try each action starting from highest Q-value
+    # Try each action starting from highest Q-value
-        for a, q_val in valid_actions:
+    for a, q_val in valid_actions:
-            s_test = list(s)
+        s_test = list(s)
-            if a == 0:  # left
+        if a == 0:  # left
-                s_test[0] -= 1
+            s_test[0] -= 1
-            elif a == 1:  # right
+        elif a == 1:  # right
-                s_test[0] += 1
+            s_test[0] += 1
-            elif a == 2:  # up
+        elif a == 2:  # up
-                s_test[1] -= 1
+            s_test[1] -= 1
-            elif a == 3:  # down
+        elif a == 3:  # down
-                s_test[1] += 1
+            s_test[1] += 1
         
-            return a    
+        return a    
 
 def calc_reward(s_new, labyrinth):
     # Reward for cookies; punish for not eating cookies
-    r = 1.0 if labyrinth[s_new[1]][s_new[0]] == "." else -1.0
+    r = 2.0 if labyrinth[s_new[1]][s_new[0]] == "." else -1.0
 
     return r