Larger layout, fixed prev score in reset

overcooked_simulator/game_content/layouts/rl.layout

-#X##
-T__#
-U__P
-#CW#
+#X###
+T___#
+#___#
+U___P
+#C#W#

overcooked_simulator/gym_env.py

@@ -243,6 +243,8 @@ class EnvGymWrapper(Env):
         info = {}
         obs = self.get_observation()
 
+        self.prev_score = 0
+
         return obs, info
 
     def get_observation(self):
@@ -288,14 +290,14 @@ def main():
 
     config = {
         "policy_type": "MlpPolicy",
-        "total_timesteps": 500_000,  # hendric sagt eher so 300_000_000 schritte
+        "total_timesteps": 100_000,  # hendric sagt eher so 300_000_000 schritte
         "env_id": "overcooked",
     }
 
     debug = True
     do_training = True
     vec_env = True
-    number_envs_parallel = 64
+    number_envs_parallel = 8
 
     model_classes = [A2C, DQN, PPO]
     model_class = model_classes[2]
@@ -307,11 +309,11 @@ def main():
 
     model_save_path = rl_agent_checkpoints / f"overcooked_{model_class.__name__}"
 
+
     if do_training:
         model = model_class(
             config["policy_type"], env, verbose=1, tensorboard_log=f"runs/{0}"
         )
-
         if debug:
             model.learn(
                 total_timesteps=config["total_timesteps"],
@@ -335,9 +337,9 @@ def main():
                 save_vecnormalize=True,
             )
             wandb_callback = WandbCallback(
-                    model_save_path=f"models/{run.id}",
-                    verbose=0,
-                )
+                model_save_path=f"models/{run.id}",
+                verbose=0,
+            )
 
             callback = CallbackList([checkpoint_callback, wandb_callback])
             model.learn(
@@ -360,6 +362,7 @@ def main():
         time.sleep(1 / 10)
         action, _states = model.predict(obs, deterministic=False)
         obs, reward, terminated, truncated, info = env.step(int(action))
+        print(reward)
         env.render()
         if terminated or truncated:
             obs, info = env.reset()