Source code for xuance.tensorflow.agents.qlearning_family.dueldqn_agent

from argparse import Namespace
from gymnasium.spaces import Space
from xuance.common import Optional, BaseCallback
from xuance.environment import DummyVecEnv, SubprocVecEnv
from xuance.tensorflow import Module
from xuance.tensorflow.utils import NormalizeFunctions, ActivationFunctions, InitializeFunctions
from xuance.tensorflow.policies import REGISTRY_Policy
from xuance.tensorflow.agents.qlearning_family.dqn_agent import DQN_Agent




class DuelDQN_Agent(DQN_Agent):
    """The implementation of DuelDQN agent.

    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[DummyVecEnv | SubprocVecEnv] = None,
            observation_space: Optional[Space] = None,
            action_space: Optional[Space] = None,
            callback: Optional[BaseCallback] = None
    ):
        super(DuelDQN_Agent, self).__init__(config, envs, observation_space, action_space, callback)

    def _build_policy(self) -> Module:
        normalize_fn = NormalizeFunctions[self.config.normalize] if hasattr(self.config, "normalize") else None
        initializer = InitializeFunctions[self.config.initialize] if hasattr(self.config, "initialize") else None
        activation = ActivationFunctions[self.config.activation]

        # build representation.
        representation = self._build_representation(self.config.representation, self.observation_space, self.config)

        # build policy.
        if self.config.policy == "Duel_Q_network":
            policy = REGISTRY_Policy["Duel_Q_network"](
                action_space=self.action_space, representation=representation, hidden_size=self.config.q_hidden_size,
                normalize=normalize_fn, initialize=initializer, activation=activation,
                use_distributed_training=self.distributed_training)
        else:
            raise AttributeError(f"{self.config.agent} currently does not support the policy named {self.config.policy}.")

        return policy