Remove seeding on reset

config.yaml

@@ -5,18 +5,18 @@ wandb:
 
 debug:
   env:
-    headless: True
+    headless: False
     stream_wrapper: False
     init_state: cut
-    max_steps: 4
+    max_steps: 1_000_000
   train:
     device: cpu
     compile: False
     compile_mode: default
-    num_envs: 10
+    num_envs: 4
     envs_per_worker: 1
-    envs_per_batch: 1
-    batch_size: 8
+    envs_per_batch: 4
+    batch_size: 16
     batch_rows: 4
     bptt_horizon: 2
     total_timesteps: 100_000_000
@@ -28,8 +28,8 @@ debug:
     env_pool: False
     log_frequency: 5000
     load_optimizer_state: False
-    swarm_frequency: 5
-    swarm_keep_pct: .8
+    swarm_frequency: 10
+    swarm_keep_pct: .1
 
 env:
   headless: True

pokemonred_puffer/environment.py

@@ -291,9 +291,9 @@ def reset(self, seed: Optional[int] = None, options: Optional[dict[str, Any]] =
                 self.moves_obtained = np.zeros(0xA5, dtype=np.uint8)
                 self.pokecenters = np.zeros(252, dtype=np.uint8)
             # lazy random seed setting
-            if not seed:
-                seed = random.randint(0, 4096)
-            self.pyboy.tick(seed, render=False)
+            # if not seed:
+            #     seed = random.randint(0, 4096)
+            #  self.pyboy.tick(seed, render=False)
         else:
             self.reset_count += 1
 

pokemonred_puffer/rewards/baseline.py

@@ -13,6 +13,7 @@
 class BaselineRewardEnv(RedGymEnv):
     def __init__(self, env_config: pufferlib.namespace, reward_config: pufferlib.namespace):
         super().__init__(env_config)
+        self.reward_config = reward_config
 
     # TODO: make the reward weights configurable
     def get_game_state_reward(self):
@@ -83,10 +84,6 @@ def get_levels_reward(self):
 
 
 class TeachCutReplicationEnv(BaselineRewardEnv):
-    def __init__(self, env_config: pufferlib.namespace, reward_config: pufferlib.namespace):
-        super().__init__(env_config)
-        self.reward_config = reward_config
-
     def get_game_state_reward(self):
         return {
             "event": self.reward_config["event"] * self.update_max_event_rew(),
@@ -117,10 +114,6 @@ def get_game_state_reward(self):
 
 
 class TeachCutReplicationEnvFork(BaselineRewardEnv):
-    def __init__(self, env_config: pufferlib.namespace, reward_config: pufferlib.namespace):
-        super().__init__(env_config)
-        self.reward_config = reward_config
-
     def get_game_state_reward(self):
         return {
             "event": self.reward_config["event"] * self.update_max_event_rew(),
@@ -172,7 +165,7 @@ def get_levels_reward(self):
             return 15 + (self.max_level_sum - 15) / 4
 
 
-class RockTunnelReplicationEnv(TeachCutReplicationEnv):
+class RockTunnelReplicationEnv(BaselineRewardEnv):
     def get_game_state_reward(self):
         return {
             "level": self.reward_config["level"] * self.get_levels_reward(),