Debugging

environments/factory/base/base_factory.py

@@ -35,7 +35,7 @@ class BaseFactory(gym.Env):
 
     @property
     def named_action_space(self):
-        return {x.identifier.value: idx for idx, x in enumerate(self._actions.values())}
+        return {x.identifier: idx for idx, x in enumerate(self._actions.values())}
 
     @property
     def observation_space(self):
@@ -287,7 +287,7 @@ class BaseFactory(gym.Env):
                 doors.tick_doors()
 
         # Finalize
-        reward, reward_info = self.build_reward_result()
+        reward, reward_info = self.build_reward_result(rewards)
 
         info.update(reward_info)
         if self._steps >= self.max_steps:
@@ -313,8 +313,8 @@ class BaseFactory(gym.Env):
             if door is not None:
                 door.use()
                 valid = c.VALID
-                self.print(f'{agent.name} just used a door {door.name}')
-                info_dict = {f'{agent.name}_door_use_{door.name}': 1}
+                self.print(f'{agent.name} just used a {door.name} at {door.pos}')
+                info_dict = {f'{agent.name}_door_use': 1}
             # When he doesn't...
             else:
                 valid = c.NOT_VALID
@@ -478,8 +478,7 @@ class BaseFactory(gym.Env):
         return oobs
 
     def get_all_tiles_with_collisions(self) -> List[Tile]:
-        tiles = [x.tile for y in self._entities for x in y if
-                 y.can_collide and not isinstance(y, WallTiles) and x.can_collide and len(x.tile.guests) > 1]
+        tiles = [x for x in self[c.FLOOR] if len(x.guests_that_can_collide) > 1]
         if False:
             tiles_with_collisions = list()
             for tile in self[c.FLOOR]:
@@ -503,11 +502,11 @@ class BaseFactory(gym.Env):
             else:
                 valid = c.NOT_VALID
                 self.print(f'{agent.name} just hit the wall at {agent.pos}.')
-                info_dict.update({f'{agent.pos}_wall_collide': 1})
+                info_dict.update({f'{agent.name}_wall_collide': 1})
         else:
             # Agent seems to be trying to Leave the level
             self.print(f'{agent.name} tried to leave the level {agent.pos}.')
-            info_dict.update({f'{agent.pos}_wall_collide': 1})
+            info_dict.update({f'{agent.name}_wall_collide': 1})
         reward_value = r.MOVEMENTS_VALID if valid else r.MOVEMENTS_FAIL
         reward = {'value': reward_value, 'reason': action.identifier, 'info': info_dict}
         return valid, reward
@@ -554,7 +553,7 @@ class BaseFactory(gym.Env):
     def additional_per_agent_rewards(self, agent) -> List[dict]:
         return []
 
-    def build_reward_result(self) -> (int, dict):
+    def build_reward_result(self, global_env_rewards: list) -> (int, dict):
         # Returns: Reward, Info
         info = defaultdict(lambda: 0.0)
 
@@ -584,12 +583,14 @@ class BaseFactory(gym.Env):
         combined_info_dict = dict(combined_info_dict)
         combined_info_dict.update(info)
 
+        global_reward_sum = sum(global_env_rewards)
         if self.individual_rewards:
             self.print(f"rewards are {comb_rewards}")
             reward = list(comb_rewards.values())
+            reward = [x + global_reward_sum for x in reward]
             return reward, combined_info_dict
         else:
-            reward = sum(comb_rewards.values())
+            reward = sum(comb_rewards.values()) + global_reward_sum
             self.print(f"reward is {reward}")
         return reward, combined_info_dict
 

environments/factory/factory_dirt.py

@@ -268,7 +268,7 @@ class DirtFactory(BaseFactory):
 
 if __name__ == '__main__':
     from environments.utility_classes import AgentRenderOptions as aro
-    render = False
+    render = True
 
     dirt_props = DirtProperties(
         initial_dirt_ratio=0.35,
@@ -293,11 +293,11 @@ if __name__ == '__main__':
     global_timings = []
     for i in range(10):
 
-        factory = DirtFactory(n_agents=2, done_at_collision=False,
+        factory = DirtFactory(n_agents=4, done_at_collision=False,
                               level_name='rooms', max_steps=1000,
                               doors_have_area=False,
                               obs_prop=obs_props, parse_doors=True,
-                              verbose=False,
+                              verbose=True,
                               mv_prop=move_props, dirt_prop=dirt_props,
                               # inject_agents=[TSPDirtAgent],
                               )
@@ -307,6 +307,7 @@ if __name__ == '__main__':
         _ = factory.observation_space
         obs_space = factory.observation_space
         obs_space_named = factory.named_observation_space
+        action_space_named = factory.named_action_space
         times = []
         for epoch in range(10):
             start_time = time.time()

environments/helpers.py

@@ -78,12 +78,12 @@ class EnvActions:
 
 class Rewards:
 
-    MOVEMENTS_VALID = -0.001
-    MOVEMENTS_FAIL  = -0.001
-    NOOP = -0.1
-    USE_DOOR_VALID = -0.001
-    USE_DOOR_FAIL  = -0.001
-    COLLISION      = -1
+    MOVEMENTS_VALID = -0.01
+    MOVEMENTS_FAIL  = -0.1
+    NOOP = -0.01
+    USE_DOOR_VALID = -0.01
+    USE_DOOR_FAIL  = -0.1
+    COLLISION      = -0.5
 
 
 m = EnvActions
@@ -120,7 +120,7 @@ class ObservationTranslator:
 
     def translate_observation(self, agent_idx: int, obs: np.ndarray):
         target_obs_space = self._per_agent_named_obs_space[agent_idx]
-        translation = [idx_space_dict['explained_idxs'] for name, idx_space_dict in target_obs_space.items()]
+        translation = [idx_space_dict for name, idx_space_dict in target_obs_space.items()]
         flat_translation = [x for y in translation for x in y]
         return np.take(obs, flat_translation, axis=1 if obs.ndim == 4 else 0)