Source code for xuance.torch.agents.policy_gradient.a2c_agent

# This is the main file for an advantage actor critic (A2C) algorithm.
# The agent random sample a batch in the replay buffer, and optimize the policy gradient and value function loss.
# This can be a first RL algorithm code for the starters.
import torch
from argparse import Namespace
from gymnasium.spaces import Space
from xuance.common import Optional, BaseCallback
from xuance.environment import DummyVecEnv, SubprocVecEnv
from xuance.torch import Module
from xuance.torch.utils import NormalizeFunctions, ActivationFunctions
from xuance.torch.policies import REGISTRY_Policy
from xuance.torch.agents import OnPolicyAgent




class A2C_Agent(OnPolicyAgent):
    """The implementation of A2C 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(A2C_Agent, self).__init__(config, envs, observation_space, action_space, callback)
        self.memory = self._build_memory()  # build memory
        self.policy = self._build_policy()  # build policy
        self.learner = self._build_learner(self.config, self.policy, self.callback)  # build learner

    def _build_policy(self) -> Module:
        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

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

        # build policy.
        if self.config.policy == "Categorical_AC":
            policy = REGISTRY_Policy["Categorical_AC"](
                action_space=self.action_space, representation=representation,
                actor_hidden_size=self.config.actor_hidden_size, critic_hidden_size=self.config.critic_hidden_size,
                normalize=normalize_fn, initialize=initializer, activation=activation, device=device,
                use_distributed_training=self.distributed_training)
        elif self.config.policy == "Gaussian_AC":
            policy = REGISTRY_Policy["Gaussian_AC"](
                action_space=self.action_space, representation=representation,
                actor_hidden_size=self.config.actor_hidden_size, critic_hidden_size=self.config.critic_hidden_size,
                normalize=normalize_fn, initialize=initializer, activation=activation, device=device,
                use_distributed_training=self.distributed_training,
                activation_action=ActivationFunctions[self.config.activation_action])
        else:
            raise AttributeError(f"A2C currently does not support the policy named {self.config.policy}.")

        return policy