DEbugging done but he does not learn

environments/factory/base/base_factory.py

@@ -181,7 +181,7 @@ class BaseFactory(gym.Env):
         # Move this in a seperate function?
         for action, agent in zip(actions, self[c.AGENT]):
             agent.clear_temp_sate()
-            action_obj = self._actions[action]
+            action_obj = self._actions[int(action)]
             if self._actions.is_moving_action(action_obj):
                 valid = self._move_or_colide(agent, action_obj)
             elif h.EnvActions.NOOP == agent.temp_action:
@@ -285,6 +285,7 @@ class BaseFactory(gym.Env):
 
         if r := self.pomdp_r:
             x, y = self._level_shape
+            self._padded_obs_cube[:] = c.SHADOWED_CELL.value
             self._padded_obs_cube[:, r:r + x, r:r + y] = self._obs_cube
             global_x, global_y = map(sum, zip(agent.pos, (r, r)))
             x0, x1 = max(0, global_x - self.pomdp_r), global_x + self.pomdp_r + 1
@@ -321,9 +322,7 @@ class BaseFactory(gym.Env):
                 light_block_map[xs, ys] = 0
             agent.temp_light_map = light_block_map
             for obs_idx in can_be_shadowed_idxs:
-                obs[obs_idx] = (obs[obs_idx] * light_block_map) - (
-                        (1 - light_block_map) * obs[0]
-                )
+                obs[obs_idx] = ((obs[obs_idx] * light_block_map) + 0.) - (1 - light_block_map)  # * obs[0])
 
             return obs
         else:
@@ -404,7 +403,7 @@ class BaseFactory(gym.Env):
                     self.print(f'{agent.name} did just use the door at {agent.pos}.')
                     info_dict.update(door_used=1)
                 else:
-                    reward -= 0.01
+                    reward -= 0.00
                     self.print(f'{agent.name} just tried to use a door at {agent.pos}, but failed.')
                     info_dict.update({f'{agent.name}_failed_action': 1})
                     info_dict.update({f'{agent.name}_failed_door_open': 1})
@@ -416,6 +415,9 @@ class BaseFactory(gym.Env):
             reward += additional_reward
             info_dict.update(additional_info_dict)
 
+            if agent.temp_collisions:
+                self.print(f't = {self._steps}\t{agent.name} has collisions with {agent.temp_collisions}')
+
             for other_agent in agent.temp_collisions:
                 info_dict.update({f'{agent.name}_vs_{other_agent.name}': 1})
 

environments/factory/base/registers.py

@@ -48,7 +48,7 @@ class Register:
         return self._register.items()
 
     def __getitem__(self, item):
-        if isinstance(item, int):
+        if isinstance(item, (int, np.int64, np.int32)):
             try:
                 return next(v for i, v in enumerate(self._register.values()) if i == item)
             except StopIteration:

environments/factory/double_task_factory.py

@@ -291,7 +291,7 @@ if __name__ == '__main__':
 
     factory = DoubleTaskFactory(item_props, n_agents=3, done_at_collision=False, frames_to_stack=0,
                                 level_name='rooms', max_steps=4000,
-                                omit_agent_slice_in_obs=True, parse_doors=True, pomdp_r=3,
+                                omit_agent_in_obs=True, parse_doors=True, pomdp_r=3,
                                 record_episodes=False, verbose=False
                                 )
 

environments/factory/simple_factory.py

@@ -18,12 +18,6 @@ from environments.utility_classes import MovementProperties
 CLEAN_UP_ACTION = h.EnvActions.CLEAN_UP
 
 
-class ObsSlice(Enum):
-    OWN = -1
-    LEVEL = c.LEVEL.value
-    AGENT = c.AGENT.value
-
-
 class DirtProperties(NamedTuple):
     clean_amount: int = 1               # How much does the robot clean with one actions.
     max_spawn_ratio: float = 0.2        # On max how much tiles does the dirt spawn in percent.
@@ -33,7 +27,6 @@ class DirtProperties(NamedTuple):
     max_global_amount: int = 20         # Max dirt amount in the whole environment.
     dirt_smear_amount: float = 0.2      # Agents smear dirt, when not cleaning up in place.
     agent_can_interact: bool = True     # Whether the agents can interact with the dirt in this environment.
-    on_obs_slice: Enum = ObsSlice.LEVEL
 
 
 class Dirt(Entity):
@@ -217,16 +210,13 @@ class SimpleFactory(BaseFactory):
         info_dict.update(dirty_tile_count=dirty_tile_count)
         info_dict.update(dirt_distribution_score=dirt_distribution_score)
 
-        if agent.temp_collisions:
-            self.print(f't = {self._steps}\t{agent.name} has collisions with {agent.temp_collisions}')
-
         if agent.temp_action == CLEAN_UP_ACTION:
             if agent.temp_valid:
                 reward += 0.5
                 self.print(f'{agent.name} did just clean up some dirt at {agent.pos}.')
                 info_dict.update(dirt_cleaned=1)
             else:
-                reward -= 0.01
+                reward -= 0.00
                 self.print(f'{agent.name} just tried to clean up some dirt at {agent.pos}, but failed.')
                 info_dict.update({f'{agent.name}_failed_action': 1})
                 info_dict.update({f'{agent.name}_failed_action': 1})
@@ -244,9 +234,9 @@ if __name__ == '__main__':
     move_props = MovementProperties(True, True, False)
 
     factory = SimpleFactory(n_agents=1, done_at_collision=False, frames_to_stack=0,
-                            level_name='rooms', max_steps=400,
-                            omit_agent_slice_in_obs=True, parse_doors=True, pomdp_r=2,
-                            record_episodes=False, verbose=False
+                            level_name='rooms', max_steps=400, combin_agent_obs=True,
+                            omit_agent_in_obs=True, parse_doors=True, pomdp_r=2,
+                            record_episodes=False, verbose=True
                             )
 
     # noinspection DuplicatedCode