Results resolved. Small ADjustments

marl_factory_grid/algorithms/marl/mappo.py

@@ -46,7 +46,7 @@ class LoopMAPPO(LoopSNAC):
         # monte carlo returns
         mc_returns = self.monte_carlo_returns(batch[nms.REWARD], batch[nms.DONE], gamma)
         mc_returns = (mc_returns - mc_returns.mean()) / (mc_returns.std() + 1e-8)  # todo: norm across agent ok?
-        advantages =  mc_returns - out[nms.CRITIC][:, :-1]
+        advantages = mc_returns - out[nms.CRITIC][:, :-1]
 
         # policy loss
         log_ap = torch.log_softmax(logits, -1)

marl_factory_grid/configs/_obs_test.yaml

@@ -0,0 +1,66 @@
+General:
+  env_seed: 69
+  individual_rewards: true
+  level_name: obs_test_map
+  pomdp_r: 0
+  verbose: True
+  tests: false
+
+Agents:
+  Wolfgang:
+    Actions:
+      - Noop
+    Observations:
+      - Walls
+      - Doors
+      - Other
+      - DirtPiles
+    Positions:
+      - (1, 3)
+
+  Soeren:
+    Actions:
+      - Noop
+    Observations:
+      - Walls
+      - Doors
+      - Other
+      - DirtPiles
+    Positions:
+      - (1, 1)
+
+  Juergen:
+    Actions:
+      - Noop
+    Observations:
+      - Walls
+      - Doors
+      - Other
+      - DirtPiles
+    Positions:
+      - (1, 2)
+
+  Walter:
+    Actions:
+      - Noop
+    Observations:
+      - Walls
+      - Doors
+      - Other
+      - DirtPiles
+    Positions:
+      - (1, 4)
+
+
+Entities:
+  DirtPiles:
+  Doors:
+
+Rules:
+  # Utilities
+  WatchCollisions:
+    done_at_collisions: false
+
+  # Done Conditions
+  DoneAtMaxStepsReached:
+    max_steps: 500

marl_factory_grid/configs/default_config.yaml

@@ -2,7 +2,7 @@ Agents:
   Wolfgang:
     Actions:
     - Noop
-    - BtryCharge
+    - Charge
     - CleanUp
     - DestAction
     - DoorUse

marl_factory_grid/configs/eight_puzzle.yaml

@@ -1,8 +1,20 @@
+General:
+  env_seed: 69
+  individual_rewards: true
+  level_name: eight_puzzle
+  pomdp_r: 0
+  verbose: True
+  tests: false
+
 Agents:
   Wolfgang:
     Actions:
-      - Noop
+      Noop:
-      - Move4
+        fail_reward: -0
+        valid_reward: 0
+      Move4:
+        fail_reward: -0.1
+        valid_reward: -.01
     Observations:
       - Other
       - Walls
@@ -35,13 +47,6 @@ Entities:
           Walter: 1
           Siggi: 1
           Dennis: 1
-General:
-  env_seed: 69
-  individual_rewards: true
-  level_name: eight_puzzle
-  pomdp_r: 3
-  verbose: True
-  tests: false
 
 Rules:
   # Utilities

marl_factory_grid/configs/narrow_corridor.yaml

@@ -81,9 +81,10 @@ Rules:
     reward_at_done: -1
     done_at_collisions: false
   # Done Conditions
-  #   Load any of the rules, to check for done conditions. 
+  #   Load any of the rules, to check for done conditions.
-  # DoneAtDestinationReachAny:
+  DoneAtDestinationReach:
-  DoneAtDestinationReachAll:
+    reward_at_done: 1
-  #  reward_at_done: 1
+    # We want to give rewards only, when all targets have been reached.
+    condition: "all"
   DoneAtMaxStepsReached:
     max_steps: 200

marl_factory_grid/environment/actions.py

@@ -1,4 +1,5 @@
 import abc
+import random
 from typing import Union
 
 from marl_factory_grid.environment import rewards as r, constants as c
@@ -13,45 +14,50 @@ class Action(abc.ABC):
         return self._identifier
 
     @abc.abstractmethod
-    def __init__(self, identifier: str):
+    def __init__(self, identifier: str, default_valid_reward: float,  default_fail_reward: float,
+                 valid_reward: float | None = None, fail_reward: float | None = None):
+        self.fail_reward = fail_reward if fail_reward is not None else default_fail_reward
+        self.valid_reward = valid_reward if valid_reward is not None else default_valid_reward
         self._identifier = identifier
 
     @abc.abstractmethod
     def do(self, entity, state) -> Union[None, ActionResult]:
-        print()
+        validity = bool(random.choice([0, 1]))
-        return
+        return self.get_result(validity, entity)
 
     def __repr__(self):
         return f'Action[{self._identifier}]'
 
+    def get_result(self, validity, entity):
+        reward = self.valid_reward if validity else self.fail_reward
+        return ActionResult(self.__class__.__name__, validity, reward=reward, entity=entity)
+
 
 class Noop(Action):
 
-    def __init__(self):
+    def __init__(self, **kwargs):
-        super().__init__(c.NOOP)
+        super().__init__(c.NOOP, r.NOOP, r.NOOP, **kwargs)
 
     def do(self, entity, *_) -> Union[None, ActionResult]:
-        return ActionResult(identifier=self._identifier, validity=c.VALID,
+        return self.get_result(c.VALID, entity)
-                            reward=r.NOOP, entity=entity)
 
 
 class Move(Action, abc.ABC):
 
     @abc.abstractmethod
-    def __init__(self, *args, **kwargs):
+    def __init__(self, identifier, **kwargs):
-        super().__init__(*args, **kwargs)
+        super().__init__(identifier, r.MOVEMENTS_VALID, r.MOVEMENTS_FAIL, **kwargs)
 
     def do(self, entity, state):
         new_pos = self._calc_new_pos(entity.pos)
         if state.check_move_validity(entity, new_pos):
             # noinspection PyUnresolvedReferences
-            move_validity = entity.move(new_pos, state)
+            move_really_was_valid = entity.move(new_pos, state)
-            reward = r.MOVEMENTS_VALID if move_validity else r.MOVEMENTS_FAIL
+            return self.get_result(move_really_was_valid, entity)
-            return ActionResult(entity=entity, identifier=self._identifier, validity=move_validity, reward=reward)
         else:  # There is no place to go, propably collision
             # This is currently handeld by the WatchCollisions rule, so that it can be switched on and off by conf.yml
             # return ActionResult(entity=entity, identifier=self._identifier, validity=c.NOT_VALID, reward=r.COLLISION)
-            return ActionResult(entity=entity, identifier=self._identifier, validity=c.NOT_VALID)
+            return self.get_result(c.NOT_VALID, entity)
 
     def _calc_new_pos(self, pos):
         x_diff, y_diff = MOVEMAP[self._identifier]
@@ -59,43 +65,43 @@ class Move(Action, abc.ABC):
 
 
 class North(Move):
-    def __init__(self, *args, **kwargs):
+    def __init__(self, **kwargs):
-        super().__init__(c.NORTH, *args, **kwargs)
+        super().__init__(c.NORTH, **kwargs)
 
 
 class NorthEast(Move):
-    def __init__(self, *args, **kwargs):
+    def __init__(self, **kwargs):
-        super().__init__(c.NORTHEAST, *args, **kwargs)
+        super().__init__(c.NORTHEAST, **kwargs)
 
 
 class East(Move):
-    def __init__(self, *args, **kwargs):
+    def __init__(self, **kwargs):
-        super().__init__(c.EAST, *args, **kwargs)
+        super().__init__(c.EAST, **kwargs)
 
 
 class SouthEast(Move):
-    def __init__(self, *args, **kwargs):
+    def __init__(self, **kwargs):
-        super().__init__(c.SOUTHEAST, *args, **kwargs)
+        super().__init__(c.SOUTHEAST, **kwargs)
 
 
 class South(Move):
-    def __init__(self, *args, **kwargs):
+    def __init__(self, **kwargs):
-        super().__init__(c.SOUTH, *args, **kwargs)
+        super().__init__(c.SOUTH, **kwargs)
 
 
 class SouthWest(Move):
-    def __init__(self, *args, **kwargs):
+    def __init__(self, **kwargs):
-        super().__init__(c.SOUTHWEST, *args, **kwargs)
+        super().__init__(c.SOUTHWEST, **kwargs)
 
 
 class West(Move):
-    def __init__(self, *args, **kwargs):
+    def __init__(self, **kwargs):
-        super().__init__(c.WEST, *args, **kwargs)
+        super().__init__(c.WEST, **kwargs)
 
 
 class NorthWest(Move):
-    def __init__(self, *args, **kwargs):
+    def __init__(self, **kwargs):
-        super().__init__(c.NORTHWEST, *args, **kwargs)
+        super().__init__(c.NORTHWEST, **kwargs)
 
 
 Move4 = [North, East, South, West]

marl_factory_grid/levels/obs_test_map.txt

@@ -0,0 +1,12 @@
+############
+#----------#
+#-#######--#
+#-#-----D--#
+#-#######--#
+#-D-----D--#
+#-#-#-#-#-##
+#----------#
+#----------#
+#----------#
+#----------#
+############

marl_factory_grid/modules/batteries/__init__.py

@@ -1,4 +1,4 @@
-from .actions import BtryCharge
+from .actions import Charge
 from .entitites import ChargePod, Battery
 from .groups import ChargePods, Batteries
 from .rules import DoneAtBatteryDischarge, BatteryDecharge

marl_factory_grid/modules/batteries/actions.py

@@ -8,10 +8,10 @@ from marl_factory_grid.environment import constants as c
 from marl_factory_grid.utils import helpers as h
 
 
-class BtryCharge(Action):
+class Charge(Action):
 
     def __init__(self):
-        super().__init__(b.ACTION_CHARGE)
+        super().__init__(b.ACTION_CHARGE, b.REWARD_CHARGE_VALID, b.Reward_CHARGE_FAIL)
 
     def do(self, entity, state) -> Union[None, ActionResult]:
         if charge_pod := h.get_first(state[b.CHARGE_PODS].by_pos(entity.pos)):
@@ -24,5 +24,4 @@ class BtryCharge(Action):
             valid = c.NOT_VALID
             state.print(f'{entity.name} failed to charged batteries at {entity.pos}.')
 
-        return ActionResult(entity=entity, identifier=self._identifier, validity=valid,
+        return self.get_result(valid, entity)
-                            reward=b.REWARD_CHARGE_VALID if valid else b.Reward_CHARGE_FAIL)

marl_factory_grid/modules/clean_up/actions.py

@@ -11,7 +11,7 @@ from marl_factory_grid.environment import constants as c
 class CleanUp(Action):
 
     def __init__(self):
-        super().__init__(d.CLEAN_UP)
+        super().__init__(d.CLEAN_UP, r.CLEAN_UP_VALID, r.CLEAN_UP_FAIL)
 
     def do(self, entity, state) -> Union[None, ActionResult]:
         if dirt := next((x for x in state.entities.pos_dict[entity.pos] if "dirt" in x.name.lower()), None):
@@ -24,13 +24,10 @@ class CleanUp(Action):
             valid = c.VALID
             print_str = f'{entity.name} did just clean up some dirt at {entity.pos}.'
             state.print(print_str)
-            reward = r.CLEAN_UP_VALID
+
-            identifier = d.CLEAN_UP
         else:
             valid = c.NOT_VALID
             print_str = f'{entity.name} just tried to clean up some dirt at {entity.pos}, but failed.'
             state.print(print_str)
-            reward = r.CLEAN_UP_FAIL
-            identifier = d.CLEAN_UP_FAIL
 
-        return ActionResult(identifier=identifier, validity=valid, reward=reward, entity=entity)
+        return self.get_result(valid, entity)

marl_factory_grid/modules/destinations/actions.py

@@ -11,7 +11,7 @@ from marl_factory_grid.environment import constants as c
 class DestAction(Action):
 
     def __init__(self):
-        super().__init__(d.DESTINATION)
+        super().__init__(d.DESTINATION, d.REWARD_WAIT_VALID, d.REWARD_WAIT_FAIL)
 
     def do(self, entity, state) -> Union[None, ActionResult]:
         if destination := state[d.DESTINATION].by_pos(entity.pos):
@@ -20,5 +20,4 @@ class DestAction(Action):
         else:
             valid = c.NOT_VALID
             state.print(f'{entity.name} just tried to do_wait_action do_wait_action at {entity.pos} but failed')
-        return ActionResult(entity=entity, identifier=self._identifier, validity=valid,
+        return self.get_result(valid, entity)
-                            reward=d.REWARD_WAIT_VALID if valid else d.REWARD_WAIT_FAIL)

marl_factory_grid/modules/doors/actions.py

@@ -1,6 +1,7 @@
 from typing import Union
 
 from marl_factory_grid.environment.actions import Action
+from marl_factory_grid.modules.doors.entitites import Door
 from marl_factory_grid.modules.doors import constants as d, rewards as r
 from marl_factory_grid.environment import constants as c
 from marl_factory_grid.utils.results import ActionResult
@@ -8,21 +9,23 @@ from marl_factory_grid.utils.results import ActionResult
 
 class DoorUse(Action):
 
-    def __init__(self):
+    def __init__(self, **kwargs):
-        super().__init__(d.ACTION_DOOR_USE)
+        super().__init__(d.ACTION_DOOR_USE, r.USE_DOOR_VALID, r.USE_DOOR_FAIL, **kwargs)
 
     def do(self, entity, state) -> Union[None, ActionResult]:
         # Check if agent really is standing on a door:
-        e = state.entities.get_entities_near_pos(entity.pos)
+        entities_close = state.entities.get_entities_near_pos(entity.pos)
-        try:
-            # Only one door opens TODO introduce loop
-            door = next(x for x in e if x.name.startswith(d.DOOR))
-            valid = door.use()
-            state.print(f'{entity.name} just used a {door.name} at {door.pos}')
-            return ActionResult(entity=entity, identifier=self._identifier, validity=valid, reward=r.USE_DOOR_VALID)
 
-        except StopIteration:
+        valid = False
-            # When he doesn't...
+        for door in [e for e in entities_close if isinstance(e, Door)]:
+            try:
+                # Will always be true, when there is at least a single door.
+                valid = door.use()
+                state.print(f'{entity.name} just used a {door.name} at {door.pos}')
+
+            except AttributeError:
+                pass
+        if not valid:
+            # When he doesn't stand necxxt to a door tell me.
             state.print(f'{entity.name} just tried to use a door at {entity.pos}, but there is none.')
-            return ActionResult(entity=entity, identifier=self._identifier,
+        return self.get_result(valid, entity)
-                                validity=c.NOT_VALID, reward=r.USE_DOOR_FAIL)

marl_factory_grid/modules/items/actions.py

@@ -9,8 +9,14 @@ from marl_factory_grid.environment import constants as c
 
 class ItemAction(Action):
 
-    def __init__(self):
+    def __init__(self, failed_dropoff_reward: float | None = None, valid_dropoff_reward: float | None = None, **kwargs):
-        super().__init__(i.ITEM_ACTION)
+        super().__init__(i.ITEM_ACTION, r.PICK_UP_FAIL, r.PICK_UP_VALID, **kwargs)
+        self.failed_drop_off_reward = failed_dropoff_reward if failed_dropoff_reward is not None else r.DROP_OFF_FAIL
+        self.valid_drop_off_reward = valid_dropoff_reward if valid_dropoff_reward is not None else r.DROP_OFF_FAIL
+
+    def get_dropoff_result(self, validity, entity):
+        reward = self.valid_drop_off_reward if validity else self.failed_drop_off_reward
+        return ActionResult(self.__name__, validity, reward=reward, entity=entity)
 
     def do(self, entity, state) -> Union[None, ActionResult]:
         inventory = state[i.INVENTORY].by_entity(entity)
@@ -23,16 +29,15 @@ class ItemAction(Action):
                 state.print(f'{entity.name} just dropped of an item at {drop_off.pos}.')
             else:
                 state.print(f'{entity.name} just tried to drop off at {entity.pos}, but failed.')
-            reward = r.DROP_OFF_VALID if valid else r.DROP_OFF_FAIL
+            return self.get_dropoff_result(valid, entity)
-            return ActionResult(entity=entity, identifier=self._identifier, validity=valid, reward=reward)
 
         elif items := state[i.ITEM].by_pos(entity.pos):
             item = items[0]
             item.change_parent_collection(inventory)
             item.set_pos(c.VALUE_NO_POS)
             state.print(f'{entity.name} just picked up an item at {entity.pos}')
-            return ActionResult(entity=entity, identifier=self._identifier, validity=c.VALID, reward=r.PICK_UP_VALID)
+            return self.get_result(c.VALID, entity)
 
         else:
             state.print(f'{entity.name} just tried to pick up an item at {entity.pos}, but failed.')
-            return ActionResult(entity=entity, identifier=self._identifier, validity=c.NOT_VALID, reward=r.PICK_UP_FAIL)
+            return self.get_result(c.NOT_VALID, entity)

marl_factory_grid/modules/machines/actions.py

@@ -12,15 +12,12 @@ from marl_factory_grid.utils import helpers as h
 class MachineAction(Action):
 
     def __init__(self):
-        super().__init__(m.MACHINE_ACTION)
+        super().__init__(m.MACHINE_ACTION, m.MAINTAIN_VALID, m.MAINTAIN_FAIL)
 
     def do(self, entity, state) -> Union[None, ActionResult]:
         if machine := h.get_first(state[m.MACHINES].by_pos(entity.pos)):
-            if valid := machine.maintain():
+            valid = machine.maintain()
-                return ActionResult(entity=entity, identifier=self._identifier, validity=valid, reward=marl_factory_grid.modules.machines.constants.MAINTAIN_VALID)
+            return self.get_result(valid, entity)
-            else:
+
-                return ActionResult(entity=entity, identifier=self._identifier, validity=valid, reward=marl_factory_grid.modules.machines.constants.MAINTAIN_FAIL)
         else:
-            return ActionResult(entity=entity, identifier=self._identifier,
+            return self.get_result(c.NOT_VALID, entity)
-                                validity=c.NOT_VALID, reward=marl_factory_grid.modules.machines.constants.MAINTAIN_FAIL
-                                )

marl_factory_grid/utils/config_parser.py

@@ -122,16 +122,28 @@ class FactoryConfigParser(object):
 
     def parse_agents_conf(self):
         parsed_agents_conf = dict()
-        base_env_actions  = self.default_actions.copy() + [c.MOVE4]
+
         for name in self.agents:
             # Actions
+            conf_actions = self.agents[name]['Actions']
             actions = list()
-            if c.DEFAULTS in self.agents[name]['Actions']:
+
-                actions.extend(self.default_actions)
+            if isinstance(conf_actions, dict):
-            actions.extend(x for x in self.agents[name]['Actions'] if x != c.DEFAULTS)
+                conf_kwargs = conf_actions.copy()
+                conf_actions = list(conf_actions.keys())
+            elif isinstance(conf_actions, list):
+                conf_kwargs = {}
+                if isinstance(conf_actions, dict):
+                    raise ValueError
+                pass
+            for action in conf_actions:
+                if action == c.DEFAULTS:
+                    actions.extend(self.default_actions)
+                else:
+                    actions.append(action)
             parsed_actions = list()
             for action in actions:
-                folder_path = MODULE_PATH if action not in base_env_actions else DEFAULT_PATH
+                folder_path = MODULE_PATH if action not in [c.MOVE8, c.NOOP, c.MOVE4] else DEFAULT_PATH
                 folder_path = Path(__file__).parent.parent / folder_path
                 try:
                     class_or_classes = locate_and_import_class(action, folder_path)
@@ -142,7 +154,7 @@ class FactoryConfigParser(object):
                 except TypeError:
                     parsed_actions.append(class_or_classes)
 
-            parsed_actions = [x() for x in parsed_actions]
+            parsed_actions = [x(**conf_kwargs.get(x, {})) for x in parsed_actions]
 
             # Observation
             observations = list()

marl_factory_grid/utils/helpers.py

@@ -224,8 +224,8 @@ def locate_and_import_class(class_name, folder_path: Union[str, PurePath] = ''):
                                   and x not in ['Entity',  'NamedTuple', 'List', 'Rule', 'Union',
                                                 'TickResult', 'ActionResult', 'Action', 'Agent',
                                                 'RenderEntity', 'TemplateRule', 'Objects', 'PositionMixin',
-                                                'IsBoundMixin', 'EnvObject', 'EnvObjects', 'Dict', 'Any'
+                                                'IsBoundMixin', 'EnvObject', 'EnvObjects', 'Dict', 'Any', 'Factory',
-                                                ]])
+                                                'Move8']])
         try:
             model_class = mod.__getattribute__(class_name)
             return model_class

test_observations.py

@@ -0,0 +1,41 @@
+from pathlib import Path
+from random import randint
+from tqdm import trange
+
+from marl_factory_grid.environment.factory import Factory
+
+from marl_factory_grid.utils.logging.envmonitor import EnvMonitor
+from marl_factory_grid.utils.logging.recorder import EnvRecorder
+from marl_factory_grid.utils.plotting.plot_single_runs import plot_single_run
+from marl_factory_grid.utils.tools import ConfigExplainer
+
+
+if __name__ == '__main__':
+    # Render at each step?
+    render = True
+
+    run_path = Path('study_out')
+
+    # Path to config File
+    path = Path('marl_factory_grid/configs/_obs_test.yaml')
+
+    # Env Init
+    factory = Factory(path)
+
+    # RL learn Loop
+    for episode in trange(10):
+        _ = factory.reset()
+        done = False
+        if render:
+            factory.render()
+        action_spaces = factory.action_space
+        while not done:
+            a = [randint(0, x.n - 1) for x in action_spaces]
+            obs_type, _, _, done, info = factory.step(a)
+            if render:
+                factory.render()
+            if done:
+                print(f'Episode {episode} done...')
+                break
+
+    print('Done!!! Goodbye....')