import torch
from argparse import Namespace
from gymnasium.spaces import Space
from xuance.common import List, Optional, MultiAgentBaseCallback
from xuance.environment import DummyVecMultiAgentEnv, SubprocVecMultiAgentEnv
from xuance.torch import Module
from xuance.torch.utils import NormalizeFunctions, ActivationFunctions
from xuance.torch.policies import REGISTRY_Policy
from xuance.torch.agents.multi_agent_rl.isac_agents import ISAC_Agents
[docs]
class MASAC_Agents(ISAC_Agents):
"""The implementation of MASAC agents.
Args:
config: the Namespace variable that provides hyperparameters and other settings.
envs: the vectorized environments.
callback: A user-defined callback function object to inject custom logic during training.
"""
def __init__(
self,
config: Namespace,
envs: Optional[DummyVecMultiAgentEnv | SubprocVecMultiAgentEnv] = None,
num_agents: Optional[int] = None,
agent_keys: Optional[List[str]] = None,
state_space: Optional[Space] = None,
observation_space: Optional[Space] = None,
action_space: Optional[Space] = None,
callback: Optional[MultiAgentBaseCallback] = None
):
super(MASAC_Agents, self).__init__(
config, envs, num_agents, agent_keys, state_space, observation_space, action_space, callback
)
def _build_policy(self) -> Module:
"""
Build representation(s) and policy(ies) for agent(s)
Returns:
policy (torch.nn.Module): A dict of policies.
"""
normalize_fn = NormalizeFunctions[self.config.normalize] if hasattr(self.config, "normalize") else None
initializer = torch.nn.init.orthogonal_
activation = ActivationFunctions[self.config.activation]
device = self.device
agent = self.config.agent
# build representations
A_representation = self._build_representation(self.config.representation, self.observation_space, self.config)
critic_in = [sum(self.observation_space[k].shape) + sum(self.action_space[k].shape) for k in self.agent_keys]
space_critic_in = {k: (sum(critic_in),) for k in self.agent_keys}
C_representation = self._build_representation(self.config.representation, space_critic_in, self.config)
# build policies
if self.config.policy == "Gaussian_MASAC_Policy":
policy = REGISTRY_Policy["Gaussian_MASAC_Policy"](
action_space=self.action_space, n_agents=self.n_agents,
actor_representation=A_representation, critic_representation=C_representation,
actor_hidden_size=self.config.actor_hidden_size,
critic_hidden_size=self.config.critic_hidden_size,
normalize=normalize_fn, initialize=initializer, activation=activation,
activation_action=ActivationFunctions[self.config.activation_action],
device=device, use_distributed_training=self.distributed_training,
use_parameter_sharing=self.use_parameter_sharing, model_keys=self.model_keys,
use_rnn=self.use_rnn, rnn=self.config.rnn if self.use_rnn else None)
self.continuous_control = True
else:
raise AttributeError(f"{agent} currently does not support the policy named {self.config.policy}.")
return policy
[docs]
def init_rnn_hidden(self, n_envs):
"""
Returns initialized hidden states of RNN if use RNN-based representations.
Parameters:
n_envs (int): The number of parallel environments.
Returns:
rnn_hidden_states: The hidden states for RNN.
"""
rnn_hidden_states = None
if self.use_rnn:
batch = n_envs * self.n_agents if self.use_parameter_sharing else n_envs
rnn_hidden_states = {k: self.policy.actor_representation[k].init_hidden(batch) for k in self.model_keys}
return rnn_hidden_states
[docs]
def init_hidden_item(self, i_env: int,
rnn_hidden: Optional[dict] = None):
"""
Returns initialized hidden states of RNN for i-th environment.
Parameters:
i_env (int): The index of environment that to be selected.
rnn_hidden (Optional[dict]): The RNN hidden states of actor representation.
"""
assert self.use_rnn is True, "This method cannot be called when self.use_rnn is False."
if self.use_parameter_sharing:
batch_index = list(range(i_env * self.n_agents, (i_env + 1) * self.n_agents))
else:
batch_index = [i_env, ]
for key in self.model_keys:
rnn_hidden[key] = self.policy.actor_representation[key].init_hidden_item(batch_index, *rnn_hidden[key])
return rnn_hidden