added max iterations

04_pacman_rl/pacman.py

@@ -166,14 +166,17 @@ def main():
         s_not_terminal = True
         labyrinth_copy = [list(row) for row in labyrinth]  # Create proper deep copy
         a = None
-        while s_not_terminal:
+        iteration = 0
+        max_iterations = 50  # Prevent infinite loops
+        while s_not_terminal and iteration < max_iterations:
+            iteration += 1
             print("s: " + str(s))
             print("q[s] before action: " + str(q[s]))
 
             a = rl.epsilon_greedy(q, s) # 0 = Left; 1 = Right ; 2 = Up ; 3 = Down
             s_new, r, labyrinth_copy = rl.take_action(s, a, labyrinth_copy)
             
-            q[s][a] += round(alpha * (r + gamma * max(q[s_new]) - q[s][a]), 2)
+            q[s][a] += round(alpha * (r + gamma * rl.max_q(q, s_new, labyrinth) - q[s][a]), 2)
             # q[s_new][a_opposite_direction[a]] += round(alpha * (r + gamma * max(q[s]) - q[s_new][a_opposite_direction[a]]), 2)
             
             s = s_new
@@ -184,8 +187,14 @@ def main():
             # Check for collisions (game over if ghost catches pacman)
             if s[0] == s[2] and s[1] == s[3]:
                 s_not_terminal = False
+                q[s][a] = 0.01
+                print("There was just a collision!!!")
+                print("s: " + str(s))
             
-            # time.sleep(0.05)
+            time.sleep(0.025)
+        
+        if iteration >= max_iterations:
+            print(f"Max iterations reached ({max_iterations}), breaking out of loop")
         
         s = (pacman.x, pacman.y, ghost.x, ghost.y) # as a tuple so the state becomes hashable
         a = rl.epsilon_greedy(q, s) # 0 = Left; 1 = Right ; 2 = Up ; 3 = Down

04_pacman_rl/reinforcement_learning.py

@@ -80,8 +80,8 @@ def epsilon_greedy(q, s, epsilon=0.2):
     a = q[s].index(q_max)
     
     return a
-    """
 
+    """
     if np.random.random() < epsilon:
         # Explore: choose random action (excluding blocked actions with Q=0)
         valid_actions = [i for i in range(len(q[s])) if q[s][i] > 0]
@@ -130,6 +130,52 @@ def bfs_distance(start, end, labyrinth):
     
     return float('inf')  # No path found
 
+def max_q(q, s_new, labyrinth):
+    """Calculate the maximum reward for all possible actions in state s_new"""
+    max_reward = float('-inf')
+    for a in range(4):
+        if q[s_new][a] > 0:  # Only consider valid (non-blocked) actions
+            s_test = list(s_new)
+            if a == 0:  # left
+                s_test[0] -= 1
+            elif a == 1:  # right
+                s_test[0] += 1
+            elif a == 2:  # up
+                s_test[1] -= 1
+            elif a == 3:  # down
+                s_test[1] += 1
+            
+            reward = calc_reward(tuple(s_test), labyrinth)
+            max_reward = max(max_reward, reward)
+    
+    return max_reward if max_reward != float('-inf') else 0.0
+
+def calc_reward(s_new, labyrinth):
+    # consider new distance between Pacman and Ghost using actual pathfinding
+    pacman_pos_new = (s_new[0], s_new[1])
+    ghost_pos = (s_new[2], s_new[3])
+
+    # distance_old = bfs_distance((s[0], s[1]), ghost_pos, labyrinth)    
+    distance_new = bfs_distance(pacman_pos_new, ghost_pos, labyrinth)
+
+    r = 0
+
+    if distance_new < 3:
+        r = -2
+    elif distance_new == 4:
+        r = 0.5
+    elif distance_new > 4:
+        r = 1
+
+    # Reward for cookies
+    r += 1.0 if labyrinth[s_new[1]][s_new[0]] == "." else -1.5
+    
+    # efficiency experiment
+    r -= 0.1
+
+    r = max(r, 0.01)
+
+    return r
 
 def take_action(s, a, labyrinth):
     labyrinth_copy = [list(row) for row in labyrinth]
@@ -144,19 +190,6 @@ def take_action(s, a, labyrinth):
     if a == 3:  # down
         s_new[1] += 1
     
-    # consider new distance between Pacman and Ghost using actual pathfinding
-    pacman_pos_new = (s_new[0], s_new[1])
-    ghost_pos = (s[2], s[3])
-    
-    distance_new = bfs_distance(pacman_pos_new, ghost_pos, labyrinth)
-    distance_old = bfs_distance((s[0], s[1]), ghost_pos, labyrinth)
-    
-    r = 0.05 * distance_new if distance_new != float('inf') else 0.0
-    
-    # Reward for eating cookies
-    r += 1.0 if labyrinth[s_new[1]][s_new[0]] == "." else -1.5
-    
-    # Ensure reward doesn't drop below 0.01
-    r = max(r, 0.01)
+    r = calc_reward(s_new, labyrinth)
     
     return tuple(s_new), r, labyrinth_copy