multi-agent-rlenv 3.5.1__tar.gz → 3.5.2__tar.gz

{multi_agent_rlenv-3.5.1 → multi_agent_rlenv-3.5.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: multi-agent-rlenv
-Version: 3.5.1
+Version: 3.5.2
 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-3.5.1 → multi_agent_rlenv-3.5.2}/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__ = "3.5.1"
+__version__ = "3.5.2"
 
 from . import models
 from .models import (

{multi_agent_rlenv-3.5.1 → multi_agent_rlenv-3.5.2}/src/marlenv/models/env.py

@@ -199,7 +199,7 @@ class MARLEnv(ABC, Generic[ActionSpaceType]):
             episode.add(step, action)
         return episode
 
-    def has_same_inouts(self, other) -> bool:
+    def has_same_inouts(self, other: "MARLEnv[ActionSpaceType]") -> bool:
         """Alias for `have_same_inouts(self, other)`."""
         if not isinstance(other, MARLEnv):
             return False

{multi_agent_rlenv-3.5.1 → multi_agent_rlenv-3.5.2}/src/marlenv/models/episode.py

@@ -66,13 +66,13 @@ class Episode:
         if target_len < self.episode_len:
             raise ValueError(f"Cannot pad episode to a smaller size: {target_len} < {self.episode_len}")
         padding_size = target_len - self.episode_len
-        obs = self.all_observations + [self.all_observations[0]] * padding_size
-        extras = self.all_extras + [self.all_extras[0]] * padding_size
-        actions = self.actions + [self.actions[0]] * padding_size
-        rewards = self.rewards + [self.rewards[0]] * padding_size
+        obs = self.all_observations + [np.zeros_like(self.all_observations[0])] * padding_size
+        extras = self.all_extras + [np.zeros_like(self.all_extras[0])] * padding_size
+        actions = self.actions + [np.zeros_like(self.actions[0])] * padding_size
+        rewards = self.rewards + [np.zeros_like(self.rewards[0])] * padding_size
         availables = self.all_available_actions + [self.all_available_actions[0]] * padding_size
-        states = self.all_states + [self.all_states[0]] * padding_size
-        states_extras = self.all_states_extras + [self.all_states_extras[0]] * padding_size
+        states = self.all_states + [np.zeros_like(self.all_states[0])] * padding_size
+        states_extras = self.all_states_extras + [np.zeros_like(self.all_states_extras[0])] * padding_size
         other = {key: value + [value[0]] * padding_size for key, value in self.other.items()}
         return Episode(
             all_observations=obs,

{multi_agent_rlenv-3.5.1 → multi_agent_rlenv-3.5.2}/src/marlenv/wrappers/potential_shaping.py

@@ -2,10 +2,15 @@ from abc import abstractmethod, ABC
 from .rlenv_wrapper import RLEnvWrapper
 from marlenv import Space, MARLEnv, Observation
 from typing import TypeVar, Optional
+import numpy as np
+import numpy.typing as npt
+
+from dataclasses import dataclass
 
 A = TypeVar("A", bound=Space)
 
 
+@dataclass
 class PotentialShaping(RLEnvWrapper[A], ABC):
     """
     Potential shaping for the Laser Learning Environment (LLE).
@@ -23,7 +28,7 @@ class PotentialShaping(RLEnvWrapper[A], ABC):
     ):
         super().__init__(env, extra_shape=extra_shape)
         self.gamma = gamma
-        self.current_potential = self.compute_potential()
+        self._current_potential = self.compute_potential()
 
     def add_extras(self, obs: Observation) -> Observation:
         """Add the extras related to potential shaping. Does nothing by default."""
@@ -31,19 +36,19 @@ class PotentialShaping(RLEnvWrapper[A], ABC):
 
     def reset(self):
         obs, state = super().reset()
-        self.current_potential = self.compute_potential()
+        self._current_potential = self.compute_potential()
         return self.add_extras(obs), state
 
     def step(self, actions):
-        phi_t = self.current_potential
+        prev_potential = self._current_potential
         step = super().step(actions)
 
-        self.current_potential = self.compute_potential()
-        shaped_reward = self.gamma * self.current_potential - phi_t
+        self._current_potential = self.compute_potential()
+        shaped_reward = self.gamma * self._current_potential - prev_potential
         step.obs = self.add_extras(step.obs)
         step.reward += shaped_reward
         return step
 
     @abstractmethod
-    def compute_potential(self) -> float:
+    def compute_potential(self) -> float | npt.NDArray[np.float32]:
         """Compute the potential of the current state of the environment."""