ctm-dqn/td3_agent.py

"""
TD3 Agent using Stable-Baselines3
适配 MultiDiscrete 动作空间的 VSL 控制
"""
import numpy as np
import torch
from stable_baselines3 import TD3
from stable_baselines3.common.noise import NormalActionNoise
import gymnasium as gym
from gymnasium import spaces


class MultiDiscreteWrapper(gym.Env):
    """将MultiDiscrete动作空间包装为连续空间供TD3使用"""

    def __init__(self, state_dim, action_dims):
        super().__init__()
        self.state_dim = state_dim
        self.action_dims = action_dims
        self.num_zones = len(action_dims)

        self.observation_space = spaces.Box(
            low=-np.inf, high=np.inf, shape=(state_dim,), dtype=np.float32
        )
        self.action_space = spaces.Box(
            low=0.0, high=1.0, shape=(self.num_zones,), dtype=np.float32
        )

    def reset(self, seed=None, options=None):
        return np.zeros(self.state_dim, dtype=np.float32), {}

    def step(self, action):
        return np.zeros(self.state_dim, dtype=np.float32), 0.0, False, False, {}


class TD3Agent:
    """TD3智能体包装器"""

    def __init__(
        self,
        state_dim: int,
        action_dims: list,
        learning_rate: float = 3e-4,
        buffer_size: int = 100000,
        learning_starts: int = 1000,
        batch_size: int = 256,
        tau: float = 0.005,
        gamma: float = 0.99,
        policy_delay: int = 2,
        device: str = "cuda",
    ):
        self.state_dim = state_dim
        self.action_dims = action_dims
        self.num_zones = len(action_dims)
        self.device = device
        self.learning_starts = learning_starts

        # 创建虚拟环境
        dummy_env = MultiDiscreteWrapper(state_dim, action_dims)

        # 动作噪声
        action_noise = NormalActionNoise(
            mean=np.zeros(self.num_zones),
            sigma=0.1 * np.ones(self.num_zones)
        )

        # 创建TD3模型
        self.model = TD3(
            "MlpPolicy",
            env=dummy_env,
            learning_rate=learning_rate,
            buffer_size=buffer_size,
            learning_starts=learning_starts,
            batch_size=batch_size,
            tau=tau,
            gamma=gamma,
            policy_delay=policy_delay,
            action_noise=action_noise,
            device=device,
            verbose=0,
        )

    def select_action(self, state: np.ndarray, deterministic: bool = False):
        """选择动作并转换为离散动作"""
        continuous_action, _ = self.model.predict(state, deterministic=deterministic)

        # 映射到离散动作
        discrete_action = np.array([
            int(cont * (self.action_dims[i] - 1) + 0.5)
            for i, cont in enumerate(continuous_action)
        ])
        discrete_action = np.clip(discrete_action, 0, [d-1 for d in self.action_dims])

        return discrete_action, 0.0, 0.0

    def store_transition(self, state, action, reward, next_state, done):
        """存储经验到replay buffer"""
        continuous_action = np.array([
            action[i] / (self.action_dims[i] - 1)
            for i in range(self.num_zones)
        ], dtype=np.float32)

        self.model.replay_buffer.add(
            state, next_state, continuous_action, reward, done, [{}]
        )

    def update(self):
        """更新策略"""
        if self.model.replay_buffer.size() < self.learning_starts:
            return {}

        # 手动更新而不是调用train()
        self.model._n_updates += 1
        gradient_steps = 1

        for _ in range(gradient_steps):
            self.model._update_learning_rate(self.model.actor.optimizer)
            self.model._update_learning_rate(self.model.critic.optimizer)

            replay_data = self.model.replay_buffer.sample(self.model.batch_size)

            with torch.no_grad():
                noise = replay_data.actions.clone().data.normal_(0, self.model.target_policy_noise)
                noise = noise.clamp(-self.model.target_noise_clip, self.model.target_noise_clip)
                next_actions = (self.model.actor_target(replay_data.next_observations) + noise).clamp(-1, 1)

                next_q_values = torch.cat(self.model.critic_target(replay_data.next_observations, next_actions), dim=1)
                next_q_values, _ = torch.min(next_q_values, dim=1, keepdim=True)
                target_q_values = replay_data.rewards + (1 - replay_data.dones) * self.model.gamma * next_q_values

            current_q_values = self.model.critic(replay_data.observations, replay_data.actions)
            critic_loss = sum(torch.nn.functional.mse_loss(current_q, target_q_values) for current_q in current_q_values)

            self.model.critic.optimizer.zero_grad()
            critic_loss.backward()
            self.model.critic.optimizer.step()

            if self.model._n_updates % self.model.policy_delay == 0:
                actor_loss = -self.model.critic.q1_forward(replay_data.observations, self.model.actor(replay_data.observations)).mean()
                self.model.actor.optimizer.zero_grad()
                actor_loss.backward()
                self.model.actor.optimizer.step()

                self.model._polyak_update(self.model.critic.parameters(), self.model.critic_target.parameters(), self.model.tau)
                self.model._polyak_update(self.model.actor.parameters(), self.model.actor_target.parameters(), self.model.tau)

        return {"actor_loss": 0.0, "critic_loss": 0.0}

    def save(self, path: str):
        """保存模型"""
        self.model.save(path)

    def load(self, path: str):
        """加载模型"""
        self.model = TD3.load(path, device=self.device)