multi-agent-rlenv 3.3.6__tar.gz → 3.3.7__tar.gz

{multi_agent_rlenv-3.3.6 → multi_agent_rlenv-3.3.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: multi-agent-rlenv
-Version: 3.3.6
+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-3.3.6 → multi_agent_rlenv-3.3.7}/src/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__ = "3.3.6"
+__version__ = "3.3.7"
 
 from . import models
 from . import wrappers

{multi_agent_rlenv-3.3.6 → multi_agent_rlenv-3.3.7}/src/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
@@ -17,8 +18,10 @@ 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
@@ -86,10 +89,12 @@ class Overcooked(MARLEnv[Sequence[int] | npt.NDArray, DiscreteActionSpace]):
     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,
@@ -185,6 +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)

{multi_agent_rlenv-3.3.6 → multi_agent_rlenv-3.3.7}/tests/test_adapters.py

@@ -183,6 +183,39 @@ def test_overcooked_obs_state():
             assert step.done
 
 
+@pytest.mark.skipif(skip_overcooked, reason="Overcooked is not installed")
+def test_overcooked_shaping():
+    from marlenv.adapters import Overcooked
+
+    UP = 0
+    DOWN = 1
+    RIGHT = 2
+    LEFT = 3
+    STAY = 4
+    INTERACT = 5
+    grid = [
+        ["X", "X", "X", "D", "X"],
+        ["X", "O", "S", "2", "X"],
+        ["X", "1", "P", " ", "X"],
+        ["X", "T", "S", " ", "X"],
+        ["X", "X", "X", "X", "X"],
+    ]
+
+    env = Overcooked.from_grid(grid, reward_shaping=True)
+    env.reset()
+    actions_rewards = [
+        ([UP, STAY], False),
+        ([INTERACT, STAY], False),
+        ([RIGHT, STAY], False),
+        ([INTERACT, STAY], True),
+    ]
+
+    for action, expected_reward in actions_rewards:
+        step = env.step(action)
+        if expected_reward:
+            assert step.reward.item() > 0
+
+
 def test_pymarl():
     LIMIT = 20
     N_AGENTS = 2