Unity 机器学习代理工具包 (ML-Agents) 是一个开源项目,它使游戏和模拟能够作为训练智能代理的环境。
您最多选择25个主题 主题必须以中文或者字母或数字开头,可以包含连字符 (-),并且长度不得超过35个字符
 
 
 
 
 

594 行
21 KiB

"""
Python Environment API for the ML-Agents toolkit
The aim of this API is to expose Agents evolving in a simulation
to perform reinforcement learning on.
This API supports multi-agent scenarios and groups similar Agents (same
observations, actions spaces and behavior) together. These groups of Agents are
identified by their BehaviorName.
For performance reasons, the data of each group of agents is processed in a
batched manner. Agents are identified by a unique AgentId identifier that
allows tracking of Agents across simulation steps. Note that there is no
guarantee that the number or order of the Agents in the state will be
consistent across simulation steps.
A simulation steps corresponds to moving the simulation forward until at least
one agent in the simulation sends its observations to Python again. Since
Agents can request decisions at different frequencies, a simulation step does
not necessarily correspond to a fixed simulation time increment.
"""
from abc import ABC, abstractmethod
from collections.abc import Mapping
from typing import (
List,
NamedTuple,
Tuple,
Optional,
Dict,
Iterator,
Any,
Mapping as MappingType,
)
from enum import IntFlag, Enum
import numpy as np
from mlagents_envs.exception import UnityActionException
AgentId = int
BehaviorName = str
class DecisionStep(NamedTuple):
"""
Contains the data a single Agent collected since the last
simulation step.
- obs is a list of numpy arrays observations collected by the agent.
- reward is a float. Corresponds to the rewards collected by the agent
since the last simulation step.
- agent_id is an int and an unique identifier for the corresponding Agent.
- action_mask is an optional list of one dimensional array of booleans.
Only available when using multi-discrete actions.
Each array corresponds to an action branch. Each array contains a mask
for each action of the branch. If true, the action is not available for
the agent during this simulation step.
"""
obs: List[np.ndarray]
reward: float
agent_id: AgentId
action_mask: Optional[List[np.ndarray]]
class DecisionSteps(Mapping):
"""
Contains the data a batch of similar Agents collected since the last
simulation step. Note that all Agents do not necessarily have new
information to send at each simulation step. Therefore, the ordering of
agents and the batch size of the DecisionSteps are not fixed across
simulation steps.
- obs is a list of numpy arrays observations collected by the batch of
agent. Each obs has one extra dimension compared to DecisionStep: the
first dimension of the array corresponds to the batch size of the batch.
- reward is a float vector of length batch size. Corresponds to the
rewards collected by each agent since the last simulation step.
- agent_id is an int vector of length batch size containing unique
identifier for the corresponding Agent. This is used to track Agents
across simulation steps.
- action_mask is an optional list of two dimensional array of booleans.
Only available when using multi-discrete actions.
Each array corresponds to an action branch. The first dimension of each
array is the batch size and the second contains a mask for each action of
the branch. If true, the action is not available for the agent during
this simulation step.
"""
def __init__(self, obs, reward, agent_id, action_mask):
self.obs: List[np.ndarray] = obs
self.reward: np.ndarray = reward
self.agent_id: np.ndarray = agent_id
self.action_mask: Optional[List[np.ndarray]] = action_mask
self._agent_id_to_index: Optional[Dict[AgentId, int]] = None
@property
def agent_id_to_index(self) -> Dict[AgentId, int]:
"""
:returns: A Dict that maps agent_id to the index of those agents in
this DecisionSteps.
"""
if self._agent_id_to_index is None:
self._agent_id_to_index = {}
for a_idx, a_id in enumerate(self.agent_id):
self._agent_id_to_index[a_id] = a_idx
return self._agent_id_to_index
def __len__(self) -> int:
return len(self.agent_id)
def __getitem__(self, agent_id: AgentId) -> DecisionStep:
"""
returns the DecisionStep for a specific agent.
:param agent_id: The id of the agent
:returns: The DecisionStep
"""
if agent_id not in self.agent_id_to_index:
raise KeyError(f"agent_id {agent_id} is not present in the DecisionSteps")
agent_index = self._agent_id_to_index[agent_id] # type: ignore
agent_obs = []
for batched_obs in self.obs:
agent_obs.append(batched_obs[agent_index])
agent_mask = None
if self.action_mask is not None:
agent_mask = []
for mask in self.action_mask:
agent_mask.append(mask[agent_index])
return DecisionStep(
obs=agent_obs,
reward=self.reward[agent_index],
agent_id=agent_id,
action_mask=agent_mask,
)
def __iter__(self) -> Iterator[Any]:
yield from self.agent_id
@staticmethod
def empty(spec: "BehaviorSpec") -> "DecisionSteps":
"""
Returns an empty DecisionSteps.
:param spec: The BehaviorSpec for the DecisionSteps
"""
obs: List[np.ndarray] = []
for sen_spec in spec.observation_specs:
obs += [np.zeros((0,) + sen_spec.shape, dtype=np.float32)]
return DecisionSteps(
obs=obs,
reward=np.zeros(0, dtype=np.float32),
agent_id=np.zeros(0, dtype=np.int32),
action_mask=None,
)
class TerminalStep(NamedTuple):
"""
Contains the data a single Agent collected when its episode ended.
- obs is a list of numpy arrays observations collected by the agent.
- reward is a float. Corresponds to the rewards collected by the agent
since the last simulation step.
- interrupted is a bool. Is true if the Agent was interrupted since the last
decision step. For example, if the Agent reached the maximum number of steps for
the episode.
- agent_id is an int and an unique identifier for the corresponding Agent.
"""
obs: List[np.ndarray]
reward: float
interrupted: bool
agent_id: AgentId
class TerminalSteps(Mapping):
"""
Contains the data a batch of Agents collected when their episode
terminated. All Agents present in the TerminalSteps have ended their
episode.
- obs is a list of numpy arrays observations collected by the batch of
agent. Each obs has one extra dimension compared to DecisionStep: the
first dimension of the array corresponds to the batch size of the batch.
- reward is a float vector of length batch size. Corresponds to the
rewards collected by each agent since the last simulation step.
- interrupted is an array of booleans of length batch size. Is true if the
associated Agent was interrupted since the last decision step. For example, if the
Agent reached the maximum number of steps for the episode.
- agent_id is an int vector of length batch size containing unique
identifier for the corresponding Agent. This is used to track Agents
across simulation steps.
"""
def __init__(self, obs, reward, interrupted, agent_id):
self.obs: List[np.ndarray] = obs
self.reward: np.ndarray = reward
self.interrupted: np.ndarray = interrupted
self.agent_id: np.ndarray = agent_id
self._agent_id_to_index: Optional[Dict[AgentId, int]] = None
@property
def agent_id_to_index(self) -> Dict[AgentId, int]:
"""
:returns: A Dict that maps agent_id to the index of those agents in
this TerminalSteps.
"""
if self._agent_id_to_index is None:
self._agent_id_to_index = {}
for a_idx, a_id in enumerate(self.agent_id):
self._agent_id_to_index[a_id] = a_idx
return self._agent_id_to_index
def __len__(self) -> int:
return len(self.agent_id)
def __getitem__(self, agent_id: AgentId) -> TerminalStep:
"""
returns the TerminalStep for a specific agent.
:param agent_id: The id of the agent
:returns: obs, reward, done, agent_id and optional action mask for a
specific agent
"""
if agent_id not in self.agent_id_to_index:
raise KeyError(f"agent_id {agent_id} is not present in the TerminalSteps")
agent_index = self._agent_id_to_index[agent_id] # type: ignore
agent_obs = []
for batched_obs in self.obs:
agent_obs.append(batched_obs[agent_index])
return TerminalStep(
obs=agent_obs,
reward=self.reward[agent_index],
interrupted=self.interrupted[agent_index],
agent_id=agent_id,
)
def __iter__(self) -> Iterator[Any]:
yield from self.agent_id
@staticmethod
def empty(spec: "BehaviorSpec") -> "TerminalSteps":
"""
Returns an empty TerminalSteps.
:param spec: The BehaviorSpec for the TerminalSteps
"""
obs: List[np.ndarray] = []
for sen_spec in spec.observation_specs:
obs += [np.zeros((0,) + sen_spec.shape, dtype=np.float32)]
return TerminalSteps(
obs=obs,
reward=np.zeros(0, dtype=np.float32),
interrupted=np.zeros(0, dtype=np.bool),
agent_id=np.zeros(0, dtype=np.int32),
)
class _ActionTupleBase(ABC):
"""
An object whose fields correspond to action data of continuous and discrete
spaces. Dimensions are of (n_agents, continuous_size) and (n_agents, discrete_size),
respectively. Note, this also holds when continuous or discrete size is
zero.
"""
def __init__(
self,
continuous: Optional[np.ndarray] = None,
discrete: Optional[np.ndarray] = None,
):
self._continuous: Optional[np.ndarray] = None
self._discrete: Optional[np.ndarray] = None
if continuous is not None:
self.add_continuous(continuous)
if discrete is not None:
self.add_discrete(discrete)
@property
def continuous(self) -> np.ndarray:
return self._continuous
@property
def discrete(self) -> np.ndarray:
return self._discrete
def add_continuous(self, continuous: np.ndarray) -> None:
if continuous.dtype != np.float32:
continuous = continuous.astype(np.float32, copy=False)
if self._discrete is None:
self._discrete = np.zeros(
(continuous.shape[0], 0), dtype=self.discrete_dtype
)
self._continuous = continuous
def add_discrete(self, discrete: np.ndarray) -> None:
if discrete.dtype != self.discrete_dtype:
discrete = discrete.astype(self.discrete_dtype, copy=False)
if self._continuous is None:
self._continuous = np.zeros((discrete.shape[0], 0), dtype=np.float32)
self._discrete = discrete
@property
@abstractmethod
def discrete_dtype(self) -> np.dtype:
pass
class ActionTuple(_ActionTupleBase):
"""
An object whose fields correspond to actions of different types.
Continuous and discrete actions are numpy arrays of type float32 and
int32, respectively and are type checked on construction.
Dimensions are of (n_agents, continuous_size) and (n_agents, discrete_size),
respectively. Note, this also holds when continuous or discrete size is
zero.
"""
@property
def discrete_dtype(self) -> np.dtype:
"""
The dtype of a discrete action.
"""
return np.int32
class ActionSpec(NamedTuple):
"""
A NamedTuple containing utility functions and information about the action spaces
for a group of Agents under the same behavior.
- num_continuous_actions is an int corresponding to the number of floats which
constitute the action.
- discrete_branch_sizes is a Tuple of int where each int corresponds to
the number of discrete actions available to the agent on an independent action branch.
"""
continuous_size: int
discrete_branches: Tuple[int, ...]
def __eq__(self, other):
return (
self.continuous_size == other.continuous_size
and self.discrete_branches == other.discrete_branches
)
def __str__(self):
return f"Continuous: {self.continuous_size}, Discrete: {self.discrete_branches}"
# For backwards compatibility
def is_discrete(self) -> bool:
"""
Returns true if this Behavior uses discrete actions
"""
return self.discrete_size > 0 and self.continuous_size == 0
# For backwards compatibility
def is_continuous(self) -> bool:
"""
Returns true if this Behavior uses continuous actions
"""
return self.discrete_size == 0 and self.continuous_size > 0
@property
def discrete_size(self) -> int:
"""
Returns a an int corresponding to the number of discrete branches.
"""
return len(self.discrete_branches)
def empty_action(self, n_agents: int) -> ActionTuple:
"""
Generates ActionTuple corresponding to an empty action (all zeros)
for a number of agents.
:param n_agents: The number of agents that will have actions generated
"""
_continuous = np.zeros((n_agents, self.continuous_size), dtype=np.float32)
_discrete = np.zeros((n_agents, self.discrete_size), dtype=np.int32)
return ActionTuple(continuous=_continuous, discrete=_discrete)
def random_action(self, n_agents: int) -> ActionTuple:
"""
Generates ActionTuple corresponding to a random action (either discrete
or continuous) for a number of agents.
:param n_agents: The number of agents that will have actions generated
"""
_continuous = np.random.uniform(
low=-1.0, high=1.0, size=(n_agents, self.continuous_size)
)
_discrete = np.zeros((n_agents, self.discrete_size), dtype=np.int32)
if self.discrete_size > 0:
_discrete = np.column_stack(
[
np.random.randint(
0,
self.discrete_branches[i], # type: ignore
size=(n_agents),
dtype=np.int32,
)
for i in range(self.discrete_size)
]
)
return ActionTuple(continuous=_continuous, discrete=_discrete)
def _validate_action(
self, actions: ActionTuple, n_agents: Optional[int], name: str
) -> ActionTuple:
"""
Validates that action has the correct action dim
for the correct number of agents and ensures the type.
"""
_expected_shape = (
(n_agents, self.continuous_size)
if n_agents is not None
else (self.continuous_size,)
)
if actions.continuous.shape != _expected_shape:
raise UnityActionException(
f"The behavior {name} needs a continuous input of dimension "
f"{_expected_shape} for (<number of agents>, <action size>) but "
f"received input of dimension {actions.continuous.shape}"
)
_expected_shape = (
(n_agents, self.discrete_size)
if n_agents is not None
else (self.discrete_size,)
)
if actions.discrete.shape != _expected_shape:
raise UnityActionException(
f"The behavior {name} needs a discrete input of dimension "
f"{_expected_shape} for (<number of agents>, <action size>) but "
f"received input of dimension {actions.discrete.shape}"
)
return actions
@staticmethod
def create_continuous(continuous_size: int) -> "ActionSpec":
"""
Creates an ActionSpec that is homogenously continuous
"""
return ActionSpec(continuous_size, ())
@staticmethod
def create_discrete(discrete_branches: Tuple[int]) -> "ActionSpec":
"""
Creates an ActionSpec that is homogenously discrete
"""
return ActionSpec(0, discrete_branches)
class DimensionProperty(IntFlag):
"""
No properties specified.
"""
UNSPECIFIED = 0
"""
No Property of the observation in that dimension. Observation can be processed with
Fully connected networks.
"""
NONE = 1
"""
Means it is suitable to do a convolution in this dimension.
"""
TRANSLATIONAL_EQUIVARIANCE = 2
"""
Means that there can be a variable number of observations in this dimension.
The observations are unordered.
"""
VARIABLE_SIZE = 4
class ObservationType(Enum):
"""
An Enum which defines the type of information carried in the observation
of the agent.
"""
# Observation information is generic.
DEFAULT = 0
# Observation contains goal information for current task.
GOAL = 1
# Observation contains reward information for current task.
REWARD = 2
# Observation contains a message from another agent.
MESSAGE = 3
class ObservationSpec(NamedTuple):
"""
A NamedTuple containing information about the observation of Agents.
- shape is a Tuple of int : It corresponds to the shape of
an observation's dimensions.
- dimension_property is a Tuple of DimensionProperties flag, one flag for each
dimension.
- observation_type is an enum of ObservationType.
"""
shape: Tuple[int, ...]
dimension_property: Tuple[DimensionProperty, ...]
observation_type: ObservationType
class BehaviorSpec(NamedTuple):
"""
A NamedTuple containing information about the observation and action
spaces for a group of Agents under the same behavior.
- observation_specs is a List of ObservationSpec NamedTuple containing
information about the information of the Agent's observations such as their shapes.
The order of the SensorSpec is the same as the order of the observations of an
agent.
- action_spec is an ActionSpec NamedTuple.
"""
observation_specs: List[ObservationSpec]
action_spec: ActionSpec
class BehaviorMapping(Mapping):
def __init__(self, specs: Dict[BehaviorName, BehaviorSpec]):
self._dict = specs
def __len__(self) -> int:
return len(self._dict)
def __getitem__(self, behavior: BehaviorName) -> BehaviorSpec:
return self._dict[behavior]
def __iter__(self) -> Iterator[Any]:
yield from self._dict
class BaseEnv(ABC):
@abstractmethod
def step(self) -> None:
"""
Signals the environment that it must move the simulation forward
by one step.
"""
@abstractmethod
def reset(self) -> None:
"""
Signals the environment that it must reset the simulation.
"""
@abstractmethod
def close(self) -> None:
"""
Signals the environment that it must close.
"""
@property
@abstractmethod
def behavior_specs(self) -> MappingType[str, BehaviorSpec]:
"""
Returns a Mapping from behavior names to behavior specs.
Agents grouped under the same behavior name have the same action and
observation specs, and are expected to behave similarly in the
environment.
Note that new keys can be added to this mapping as new policies are instantiated.
"""
@abstractmethod
def set_actions(self, behavior_name: BehaviorName, action: ActionTuple) -> None:
"""
Sets the action for all of the agents in the simulation for the next
step. The Actions must be in the same order as the order received in
the DecisionSteps.
:param behavior_name: The name of the behavior the agents are part of
:param action: ActionTuple tuple of continuous and/or discrete action.
Actions are np.arrays with dimensions (n_agents, continuous_size) and
(n_agents, discrete_size), respectively.
"""
@abstractmethod
def set_action_for_agent(
self, behavior_name: BehaviorName, agent_id: AgentId, action: ActionTuple
) -> None:
"""
Sets the action for one of the agents in the simulation for the next
step.
:param behavior_name: The name of the behavior the agent is part of
:param agent_id: The id of the agent the action is set for
:param action: ActionTuple tuple of continuous and/or discrete action
Actions are np.arrays with dimensions (1, continuous_size) and
(1, discrete_size), respectively. Note, this initial dimensions of 1 is because
this action is meant for a single agent.
"""
@abstractmethod
def get_steps(
self, behavior_name: BehaviorName
) -> Tuple[DecisionSteps, TerminalSteps]:
"""
Retrieves the steps of the agents that requested a step in the
simulation.
:param behavior_name: The name of the behavior the agents are part of
:return: A tuple containing :
- A DecisionSteps NamedTuple containing the observations,
the rewards, the agent ids and the action masks for the Agents
of the specified behavior. These Agents need an action this step.
- A TerminalSteps NamedTuple containing the observations,
rewards, agent ids and interrupted flags of the agents that had their
episode terminated last step.
"""