multi-agent-rlenv 3.3.5__py3-none-any.whl → 3.3.7__py3-none-any.whl

marlenv/__init__.py

@@ -62,7 +62,7 @@ print(env.extras_shape) # (1, )
 If you want to create a new environment, you can simply create a class that inherits from `MARLEnv`. If you want to create a wrapper around an existing `MARLEnv`, you probably want to subclass `RLEnvWrapper` which implements a default behaviour for every method.
 """
 
-__version__ = "3.3.5"
+__version__ = "3.3.7"
 
 from . import models
 from . import wrappers

marlenv/adapters/__init__.py

@@ -20,7 +20,7 @@ if find_spec("smac") is not None:
     HAS_SMAC = True
 
 HAS_OVERCOOKED = False
-if find_spec("overcooked_ai_py.mdp") is not None:
+if find_spec("overcooked_ai_py") is not None and find_spec("overcooked_ai_py.mdp") is not None:
     import numpy
 
     # Overcooked assumes a version of numpy <2.0 where np.Inf is available.

marlenv/adapters/gym_adapter.py

@@ -42,26 +42,26 @@ class Gym(MARLEnv[Sequence | npt.NDArray, ActionSpace]):
             case other:
                 raise NotImplementedError(f"Action space {other} not supported")
         super().__init__(space, env.observation_space.shape, (1,))
-        self.env = env
-        if self.env.unwrapped.spec is not None:
-            self.name = self.env.unwrapped.spec.id
+        self._gym_env = env
+        if self._gym_env.unwrapped.spec is not None:
+            self.name = self._gym_env.unwrapped.spec.id
         else:
             self.name = "gym-no-id"
-        self.last_obs = None
+        self._last_obs = None
 
     def get_observation(self):
-        if self.last_obs is None:
+        if self._last_obs is None:
             raise ValueError("No observation available. Call reset() first.")
-        return self.last_obs
+        return self._last_obs
 
     def step(self, actions):
-        obs, reward, done, truncated, info = self.env.step(list(actions)[0])
-        self.last_obs = Observation(
+        obs, reward, done, truncated, info = self._gym_env.step(list(actions)[0])
+        self._last_obs = Observation(
             np.array([obs], dtype=np.float32),
             self.available_actions(),
         )
         return Step(
-            self.last_obs,
+            self._last_obs,
             self.get_state(),
             np.array([reward]),
             done,
@@ -73,18 +73,18 @@ class Gym(MARLEnv[Sequence | npt.NDArray, ActionSpace]):
         return State(np.zeros(1, dtype=np.float32))
 
     def reset(self):
-        obs_data, _info = self.env.reset()
-        self.last_obs = Observation(
+        obs_data, _info = self._gym_env.reset()
+        self._last_obs = Observation(
             np.array([obs_data], dtype=np.float32),
             self.available_actions(),
         )
-        return self.last_obs, self.get_state()
+        return self._last_obs, self.get_state()
 
     def get_image(self):
-        image = np.array(self.env.render())
+        image = np.array(self._gym_env.render())
         if sys.platform in ("linux", "linux2"):
             image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
         return image
 
     def seed(self, seed_value: int):
-        self.env.reset(seed=seed_value)
+        self._gym_env.reset(seed=seed_value)

marlenv/adapters/overcooked_adapter.py

@@ -2,6 +2,7 @@ import sys
 from dataclasses import dataclass
 from typing import Literal, Sequence
 from copy import deepcopy
+from time import time
 
 import cv2
 import numpy as np
@@ -17,14 +18,17 @@ from overcooked_ai_py.visualization.state_visualizer import StateVisualizer
 @dataclass
 class Overcooked(MARLEnv[Sequence[int] | npt.NDArray, DiscreteActionSpace]):
     horizon: int
+    reward_shaping: bool
 
-    def __init__(self, oenv: OvercookedEnv):
+    def __init__(self, oenv: OvercookedEnv, reward_shaping: bool = True):
+        self.reward_shaping = reward_shaping
         self._oenv = oenv
         assert isinstance(oenv.mdp, OvercookedGridworld)
         self._mdp = oenv.mdp
         self._visualizer = StateVisualizer()
-        shape = tuple(int(s) for s in self._mdp.get_lossless_state_encoding_shape())
-        shape = (shape[2], shape[0], shape[1])
+        width, height, layers = tuple(self._mdp.lossless_state_encoding_shape)
+        # -1 because we extract the "urgent" layer to the extras
+        shape = (int(layers - 1), int(width), int(height))
         super().__init__(
             action_space=DiscreteActionSpace(
                 n_agents=self._mdp.num_players,
@@ -32,10 +36,10 @@ class Overcooked(MARLEnv[Sequence[int] | npt.NDArray, DiscreteActionSpace]):
                 action_names=[Action.ACTION_TO_CHAR[a] for a in Action.ALL_ACTIONS],
             ),
             observation_shape=shape,
-            extras_shape=(1,),
-            extras_meanings=["timestep"],
+            extras_shape=(2,),
+            extras_meanings=["timestep", "urgent"],
             state_shape=shape,
-            state_extra_shape=(1,),
+            state_extra_shape=(2,),
             reward_space=ContinuousSpace.from_shape(1),
         )
         self.horizon = int(self._oenv.horizon)
@@ -53,19 +57,25 @@ class Overcooked(MARLEnv[Sequence[int] | npt.NDArray, DiscreteActionSpace]):
         return self.state.timestep
 
     def _state_data(self):
-        state = np.array(self._mdp.lossless_state_encoding(self.state), dtype=np.float32)
+        players_layers = self._mdp.lossless_state_encoding(self.state)
+        state = np.array(players_layers, dtype=np.float32)
         # Use axes (agents, channels, height, width) instead of (agents, height, width, channels)
         state = np.transpose(state, (0, 3, 1, 2))
-        return state
+        # The last last layer is for "urgency", put it in the extras
+        urgency = float(np.all(state[:, -1]))
+        state = state[:, :-1]
+        return state, urgency
 
     def get_state(self):
-        return State(self._state_data()[0], np.array([self.time_step / self.horizon], dtype=np.float32))
+        data, is_urgent = self._state_data()
+        return State(data[0], np.array([self.time_step / self.horizon, is_urgent], dtype=np.float32))
 
     def get_observation(self) -> Observation:
+        data, is_urgent = self._state_data()
         return Observation(
-            data=self._state_data(),
+            data=data,
             available_actions=self.available_actions(),
-            extras=np.array([[self.time_step / self.horizon]] * self.n_agents, dtype=np.float32),
+            extras=np.array([[self.time_step / self.horizon, is_urgent]] * self.n_agents, dtype=np.float32),
         )
 
     def available_actions(self):
@@ -74,15 +84,17 @@ class Overcooked(MARLEnv[Sequence[int] | npt.NDArray, DiscreteActionSpace]):
         for agent_num, agent_actions in enumerate(actions):
             for action in agent_actions:
                 available_actions[agent_num, Action.ACTION_TO_INDEX[action]] = True
-        return np.array(available_actions)
+        return np.array(available_actions, dtype=np.bool)
 
     def step(self, actions: Sequence[int] | npt.NDArray[np.int32 | np.int64]) -> Step:
         actions = [Action.ALL_ACTIONS[a] for a in actions]
         _, reward, done, info = self._oenv.step(actions, display_phi=True)
+        if self.reward_shaping:
+            reward += sum(info["shaped_r_by_agent"])
         return Step(
             obs=self.get_observation(),
             state=self.get_state(),
-            reward=np.array([reward]),
+            reward=np.array([reward], dtype=np.float32),
             done=done,
             truncated=False,
             info=info,
@@ -178,6 +190,32 @@ class Overcooked(MARLEnv[Sequence[int] | npt.NDArray, DiscreteActionSpace]):
             "you_shall_not_pass",
         ],
         horizon: int = 400,
+        reward_shaping: bool = True,
     ):
         mdp = OvercookedGridworld.from_layout_name(layout)
-        return Overcooked(OvercookedEnv.from_mdp(mdp, horizon=horizon))
+        return Overcooked(OvercookedEnv.from_mdp(mdp, horizon=horizon), reward_shaping=reward_shaping)
+
+    @staticmethod
+    def from_grid(
+        grid: Sequence[Sequence[Literal["S", "P", "X", "O", "D", "T", "1", "2", " "] | str]],
+        horizon: int = 400,
+        reward_shaping: bool = True,
+    ):
+        """
+        Create an Overcooked environment from a grid layout where
+        - S is a serving location
+        - P is a cooking pot
+        - X is a counter
+        - O is an onion dispenser
+        - D is a dish dispenser
+        - T is a tomato dispenser
+        - 1 is a player 1 starting location
+        - 2 is a player 2 starting location
+        - ' ' is a walkable space
+        """
+        # It is necessary to add an explicit layout name because Overcooked saves some files under this
+        # name. By default the name is a concatenation of the grid elements, which may include characters
+        # such as white spaces, pipes ('|') and square brackets ('[' and ']') that are invalid Windows file paths.
+        layout_name = str(time())
+        mdp = OvercookedGridworld.from_grid(grid, base_layout_params={"layout_name": layout_name})
+        return Overcooked(OvercookedEnv.from_mdp(mdp, horizon=horizon), reward_shaping=reward_shaping)

marlenv/adapters/pettingzoo_adapter.py

@@ -32,7 +32,7 @@ class PettingZoo(MARLEnv[npt.NDArray, ActionSpace]):
         obs_space = env.observation_space(env.possible_agents[0])
         if obs_space.shape is None:
             raise NotImplementedError("Only discrete observation spaces are supported")
-        self._env = env
+        self._pz_env = env
         env.reset()
         super().__init__(space, obs_space.shape, self.get_state().shape)
         self.agents = env.possible_agents
@@ -40,13 +40,13 @@ class PettingZoo(MARLEnv[npt.NDArray, ActionSpace]):
 
     def get_state(self):
         try:
-            return self._env.state()
+            return self._pz_env.state()
         except NotImplementedError:
             return np.array([0])
 
     def step(self, actions: npt.NDArray | Sequence):
         action_dict = dict(zip(self.agents, actions))
-        obs, reward, term, trunc, info = self._env.step(action_dict)
+        obs, reward, term, trunc, info = self._pz_env.step(action_dict)
         obs_data = np.array([v for v in obs.values()])
         reward = np.sum([r for r in reward.values()], keepdims=True)
         self.last_observation = Observation(obs_data, self.available_actions())
@@ -54,7 +54,7 @@ class PettingZoo(MARLEnv[npt.NDArray, ActionSpace]):
         return Step(self.last_observation, state, reward, any(term.values()), any(trunc.values()), info)
 
     def reset(self):
-        obs = self._env.reset()[0]
+        obs = self._pz_env.reset()[0]
         obs_data = np.array([v for v in obs.values()])
         self.last_observation = Observation(obs_data, self.available_actions(), self.get_state())
         return self.last_observation
@@ -65,7 +65,7 @@ class PettingZoo(MARLEnv[npt.NDArray, ActionSpace]):
         return self.last_observation
 
     def seed(self, seed_value: int):
-        self._env.reset(seed=seed_value)
+        self._pz_env.reset(seed=seed_value)
 
     def render(self, *_):
-        return self._env.render()
+        return self._pz_env.render()

marlenv/env_builder.py

@@ -127,7 +127,7 @@ class Builder(Generic[A, AS]):
         from marlenv.models import DiscreteActionSpace
 
         assert isinstance(self._env.action_space, DiscreteActionSpace)
-        self._env = wrappers.Centralised(self._env)  # type: ignore
+        self._env = wrappers.Centralized(self._env)  # type: ignore
         return self
 
     def record(

marlenv/wrappers/__init__.py

@@ -7,7 +7,7 @@ from .paddings import PadObservations, PadExtras
 from .penalty_wrapper import TimePenalty
 from .available_actions_wrapper import AvailableActions
 from .blind_wrapper import Blind
-from .centralised import Centralised
+from .centralised import Centralized
 from .available_actions_mask import AvailableActionsMask
 from .delayed_rewards import DelayedReward
 
@@ -24,6 +24,6 @@ __all__ = [
     "TimePenalty",
     "AvailableActions",
     "Blind",
-    "Centralised",
+    "Centralized",
     "DelayedReward",
 ]

marlenv/wrappers/centralised.py

@@ -14,7 +14,7 @@ A = TypeVar("A", bound=npt.NDArray | Sequence[int] | Sequence[Sequence[float]])
 
 
 @dataclass
-class Centralised(RLEnvWrapper[A, DiscreteActionSpace]):
+class Centralized(RLEnvWrapper[A, DiscreteActionSpace]):
     joint_action_space: ActionSpace
 
     def __init__(self, env: MARLEnv[A, DiscreteActionSpace]):

{multi_agent_rlenv-3.3.5.dist-info → multi_agent_rlenv-3.3.7.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: multi-agent-rlenv
-Version: 3.3.5
+Version: 3.3.7
 Summary: A strongly typed Multi-Agent Reinforcement Learning framework
 Project-URL: repository, https://github.com/yamoling/multi-agent-rlenv
 Author-email: Yannick Molinghen <yannick.molinghen@ulb.be>

{multi_agent_rlenv-3.3.5.dist-info → multi_agent_rlenv-3.3.7.dist-info}/RECORD

@@ -1,13 +1,13 @@
-marlenv/__init__.py,sha256=XbptMcX24_x5Tk4G6ff6hjI-FkE9MKHLULtCrmvEKl8,3741
-marlenv/env_builder.py,sha256=_rdwcWRqnHP7i4M4Oje1Y2nrEBKH9EzTpqOuw_PNUyw,5560
+marlenv/__init__.py,sha256=u27-QdgKv_1k3uR0oCBN7wcX2jRPmpICbJz1SaZ-f-A,3741
+marlenv/env_builder.py,sha256=s_lQANqP3iNc8nmcr3CanRVsExnn9qh0ihh4lFr0c4c,5560
 marlenv/env_pool.py,sha256=R3WIrnQ5Zvff4HR1ecfkDmuO2zl7v1ywQ0K2_nvWFzs,1070
 marlenv/exceptions.py,sha256=gJUC_2rVAvOfK_ypVFc7Myh-pIfSU3To38VBVS_0rZA,1179
 marlenv/mock_env.py,sha256=qB0fYFIfbopJf7Va8kCeVI5vsOy1-2JdEYe9gdV1Ruw,4761
 marlenv/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-marlenv/adapters/__init__.py,sha256=NEmuHPWz4SGQcgF7QuIeA0QaXK141JoYco-7mqj9Ghk,883
-marlenv/adapters/gym_adapter.py,sha256=Vx6ZrYI7kiNlJODmqyjXu9WCdbCr6trcMNot0pvYD74,2864
-marlenv/adapters/overcooked_adapter.py,sha256=B6umkRG6POUQNkeHgdIVuYRv8WjIA-3S1EVP6uaKwGY,6458
-marlenv/adapters/pettingzoo_adapter.py,sha256=9rwSc_b7qV3ChtEIevOkJvtIp7WoY3CVnu6L9DxlMB4,2852
+marlenv/adapters/__init__.py,sha256=rWiqQOqTx3kVL5ZkPo3rkczrlQBBhQbU55zGI26SEeY,929
+marlenv/adapters/gym_adapter.py,sha256=6CBEjANViTJBTUBmtVyrhJrzjBJxNs_4hmMnXXG2mkU,2906
+marlenv/adapters/overcooked_adapter.py,sha256=Ehwwha_gh9wsQWBVLvwKYR_P6WUco-W2LoxumVjXSPQ,8289
+marlenv/adapters/pettingzoo_adapter.py,sha256=4F1au6uctsqRhGfcZOeDRH-8hmrFXnA5xH1Z1Pnek3s,2870
 marlenv/adapters/pymarl_adapter.py,sha256=x__E90XpFbfSWhnBHtkcD6WYkmKki1LByNbUFoDBUcg,3416
 marlenv/adapters/smac_adapter.py,sha256=fOfKo1hL4ioKtM5qQGcwtfdkdwUEACjAZqaGmkoQUcU,8373
 marlenv/models/__init__.py,sha256=9M-rnj94nsdyO4zm_VEtyYBmde3iD2_eIY4bMB-IqCo,555
@@ -18,12 +18,12 @@ marlenv/models/spaces.py,sha256=pw8Sum_fHBkR-lyfTqUij4azMCNm8oBZrYZe4WVR7rA,7652
 marlenv/models/state.py,sha256=958PXTHadi3gtRnhGgcGtqBnF44R11kdcx62NN2gwxA,1717
 marlenv/models/step.py,sha256=LKGAV2Cu-k9Gz1hwrfvGx51l8axtQRqDE9WVL5r2A1Q,3037
 marlenv/models/transition.py,sha256=2vvuhSSq911weCXio9nuyfsLVh_7ORSU_znOqpLLdLg,5107
-marlenv/wrappers/__init__.py,sha256=P7YCK1KYJvE6BAlH--nOW9PSlrohhuw-1wlfgCTOl9U,808
+marlenv/wrappers/__init__.py,sha256=wl23NUYcl0vPJb2QLpe4Xj8ZocUIOarAZX8CgWqdSQE,808
 marlenv/wrappers/agent_id_wrapper.py,sha256=oTIAYxKD1JtHfrZN43mf-3e8pxjd0nxm07vxs3BfrGY,1187
 marlenv/wrappers/available_actions_mask.py,sha256=JoCJ9eqHlkY8wfY-oaceEi8yp1Efs1iK6IO2Ibf9oZA,1468
 marlenv/wrappers/available_actions_wrapper.py,sha256=9UTwP3LXvncBITJeQnEqwiP_lj-ipULACkGs-2QbMrI,1026
 marlenv/wrappers/blind_wrapper.py,sha256=YEayRf_dclhzx6LXsasZ-IM7C71kyPb1gV0pHYYfjig,857
-marlenv/wrappers/centralised.py,sha256=J4hOMRT2fit936LifANNJtP7UbBEb_xIyF4VL9-fZGw,3226
+marlenv/wrappers/centralised.py,sha256=E2Saezwfw9uWX4R9jDgH3iQLW8c_h2Y1TdtmUi9nIvs,3226
 marlenv/wrappers/delayed_rewards.py,sha256=6oGJe-L_gGI-pQMResbkjsMDvXpni2SQvnTQ6wsZqGo,1170
 marlenv/wrappers/last_action_wrapper.py,sha256=u7a3Da5sg_gMrwZ3SE7PAwt2m9xSYYDKjngQyOmcJ74,2886
 marlenv/wrappers/paddings.py,sha256=VQOF4zaP61R74tQ4XTTT-FkK2QSy31AukICnqCy6zB0,2119
@@ -31,7 +31,7 @@ marlenv/wrappers/penalty_wrapper.py,sha256=v4_H8OEN2-yujLzRb6P7W7KwmXHtjAFsxcdp3
 marlenv/wrappers/rlenv_wrapper.py,sha256=C2XekgBIM4x3Wa2Mtsn7rihRD4ymC2hORI473Af0sfw,2962
 marlenv/wrappers/time_limit.py,sha256=CDIMMJPMyIDHSFxUJaC7nb7Kd86-07NgZeFhrpZm82o,3985
 marlenv/wrappers/video_recorder.py,sha256=d5AFu6qHqby9mOcBsYWYPxAPiK1vtnfMYdZ81AnCekI,2624
-multi_agent_rlenv-3.3.5.dist-info/METADATA,sha256=Yts2uzrt6q7VFqf6QNjwtYsuDlTHWDcQHKxKkFzptFg,4897
-multi_agent_rlenv-3.3.5.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-multi_agent_rlenv-3.3.5.dist-info/licenses/LICENSE,sha256=_eeiGVoIJ7kYt6l1zbIvSBQppTnw0mjnYk1lQ4FxEjE,1074
-multi_agent_rlenv-3.3.5.dist-info/RECORD,,
+multi_agent_rlenv-3.3.7.dist-info/METADATA,sha256=zAUWp4QbbEnxN7EVkJfJAXKYEjAcfUIKaogM43fAjR8,4897
+multi_agent_rlenv-3.3.7.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+multi_agent_rlenv-3.3.7.dist-info/licenses/LICENSE,sha256=_eeiGVoIJ7kYt6l1zbIvSBQppTnw0mjnYk1lQ4FxEjE,1074
+multi_agent_rlenv-3.3.7.dist-info/RECORD,,