PyTorch 深度学习实战(15):Twin Delayed DDPG (TD3) 算法

liuian 2025-05-08 19:41 84 浏览

在上一篇文章中，我们介绍了 Deep Deterministic Policy Gradient (DDPG) 算法，并使用它解决了 Pendulum 问题。本文将深入探讨 Twin Delayed DDPG (TD3) 算法，这是一种改进的 DDPG 算法，能够有效解决 DDPG 中的过估计问题。我们将使用 PyTorch 实现 TD3 算法，并应用于经典的 Pendulum 问题。

一、TD3 算法基础

TD3 是 DDPG 的改进版本，通过引入以下三个关键技术来解决 DDPG 中的过估计问题：

1.双重 Critic 网络：

使用两个 Critic 网络来估计 Q 值，从而减少过估计问题。

2.延迟更新：

延迟 Actor 网络的更新，确保 Critic 网络更稳定地收敛。

3.目标策略平滑：

在目标动作中加入噪声，从而减少 Critic 网络的过拟合。

1. TD3 的核心思想

双重 Critic 网络：

使用两个 Critic 网络来估计 Q 值，取两者中的较小值作为目标 Q 值，从而减少过估计。

延迟更新：

每更新 Critic 网络多次，才更新一次 Actor 网络，确保 Critic 网络更稳定地收敛。

目标策略平滑：

在目标动作中加入噪声，从而减少 Critic 网络的过拟合。

2. TD3 的优势

减少过估计：

通过双重 Critic 网络和目标策略平滑，TD3 能够有效减少 Q 值的过估计。

训练稳定：

延迟更新策略确保 Critic 网络更稳定地收敛。

适用于连续动作空间：

TD3 能够直接输出连续动作，适用于机器人控制、自动驾驶等任务。

3. TD3 的算法流程

使用当前策略采样一批数据。
使用目标网络计算目标 Q 值。
更新 Critic 网络以最小化 Q 值的误差。
延迟更新 Actor 网络以最大化 Q 值。
更新目标网络。
重复上述过程，直到策略收敛。

二、Pendulum 问题实战

我们将使用 PyTorch 实现 TD3 算法，并应用于 Pendulum 问题。目标是控制摆杆使其保持直立。

1. 问题描述

Pendulum 环境的状态空间包括摆杆的角度和角速度。动作空间是一个连续的扭矩值，范围在 -2,2 之间。智能体每保持摆杆直立一步，就会获得一个负的奖励，目标是最大化累积奖励。

2. 实现步骤

安装并导入必要的库。
定义 Actor 网络和 Critic 网络。
定义 TD3 训练过程。
测试模型并评估性能。

3. 代码实现

以下是完整的代码实现：

import gym
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import numpy as np
import random
from collections import deque
import matplotlib.pyplot as plt

plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"使用设备: {device}")

env = gym.make('Pendulum-v1')
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
max_action = float(env.action_space.high[0])

SEED = 42
torch.manual_seed(SEED)
np.random.seed(SEED)
random.seed(SEED)


# 改进的 Actor 网络（增加层归一化）
class Actor(nn.Module):
    def __init__(self, state_dim, action_dim, max_action):
        super(Actor, self).__init__()
        self.l1 = nn.Linear(state_dim, 256)
        self.ln1 = nn.LayerNorm(256)
        self.l2 = nn.Linear(256, 256)
        self.ln2 = nn.LayerNorm(256)
        self.l3 = nn.Linear(256, action_dim)
        self.max_action = max_action

    def forward(self, x):
        x = F.relu(self.ln1(self.l1(x)))
        x = F.relu(self.ln2(self.l2(x)))
        x = torch.tanh(self.l3(x)) * self.max_action
        return x


# 改进的 Critic 网络（增加层归一化）
class Critic(nn.Module):
    def __init__(self, state_dim, action_dim):
        super(Critic, self).__init__()
        self.l1 = nn.Linear(state_dim + action_dim, 256)
        self.ln1 = nn.LayerNorm(256)
        self.l2 = nn.Linear(256, 256)
        self.ln2 = nn.LayerNorm(256)
        self.l3 = nn.Linear(256, 1)

    def forward(self, x, u):
        x = F.relu(self.ln1(self.l1(torch.cat([x, u], 1))))
        x = F.relu(self.ln2(self.l2(x)))
        x = self.l3(x)
        return x


class TD3:
    def __init__(self, state_dim, action_dim, max_action):
        self.actor = Actor(state_dim, action_dim, max_action).to(device)
        self.actor_target = Actor(state_dim, action_dim, max_action).to(device)
        self.actor_target.load_state_dict(self.actor.state_dict())
        self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=3e-4)

        self.critic1 = Critic(state_dim, action_dim).to(device)
        self.critic2 = Critic(state_dim, action_dim).to(device)
        self.critic1_target = Critic(state_dim, action_dim).to(device)
        self.critic2_target = Critic(state_dim, action_dim).to(device)
        self.critic1_target.load_state_dict(self.critic1.state_dict())
        self.critic2_target.load_state_dict(self.critic2.state_dict())
        self.critic1_optimizer = optim.Adam(self.critic1.parameters(), lr=3e-4)
        self.critic2_optimizer = optim.Adam(self.critic2.parameters(), lr=3e-4)

        self.max_action = max_action
        self.replay_buffer = deque(maxlen=1000000)
        self.batch_size = 256
        self.gamma = 0.99
        self.tau = 0.005
        self.policy_noise = 0.2
        self.noise_clip = 0.5
        self.policy_freq = 2
        self.total_it = 0
        self.exploration_noise = 0.1  # 新增探索噪声

    def select_action(self, state, add_noise=True):
        state = torch.FloatTensor(state.reshape(1, -1)).to(device)
        action = self.actor(state).cpu().data.numpy().flatten()
        if add_noise:
            noise = np.random.normal(0, self.exploration_noise, size=action_dim)
            action = (action + noise).clip(-self.max_action, self.max_action)
        return action

    def train(self):
        if len(self.replay_buffer) < self.batch_size:
            return

        self.total_it += 1

        batch = random.sample(self.replay_buffer, self.batch_size)
        state = torch.FloatTensor(np.array([t[0] for t in batch])).to(device)
        action = torch.FloatTensor(np.array([t[1] for t in batch])).to(device)
        reward = torch.FloatTensor(np.array([t[2] for t in batch])).reshape(-1, 1).to(device) / 10.0  # 奖励缩放
        next_state = torch.FloatTensor(np.array([t[3] for t in batch])).to(device)
        done = torch.FloatTensor(np.array([t[4] for t in batch])).reshape(-1, 1).to(device)

        with torch.no_grad():
            noise = (torch.randn_like(action) * self.policy_noise).clamp(-self.noise_clip, self.noise_clip)
            next_action = (self.actor_target(next_state) + noise).clamp(-self.max_action, self.max_action)
            target_Q1 = self.critic1_target(next_state, next_action)
            target_Q2 = self.critic2_target(next_state, next_action)
            target_Q = torch.min(target_Q1, target_Q2)
            target_Q = reward + (1 - done) * self.gamma * target_Q

        current_Q1 = self.critic1(state, action)
        current_Q2 = self.critic2(state, action)
        critic1_loss = F.mse_loss(current_Q1, target_Q)
        critic2_loss = F.mse_loss(current_Q2, target_Q)
        self.critic1_optimizer.zero_grad()
        critic1_loss.backward()
        torch.nn.utils.clip_grad_norm_(self.critic1.parameters(), 1.0)  # 梯度裁剪
        self.critic1_optimizer.step()
        self.critic2_optimizer.zero_grad()
        critic2_loss.backward()
        torch.nn.utils.clip_grad_norm_(self.critic2.parameters(), 1.0)
        self.critic2_optimizer.step()

        if self.total_it % self.policy_freq == 0:
            actor_loss = -self.critic1(state, self.actor(state)).mean()
            self.actor_optimizer.zero_grad()
            actor_loss.backward()
            torch.nn.utils.clip_grad_norm_(self.actor.parameters(), 1.0)
            self.actor_optimizer.step()

            for param, target_param in zip(self.critic1.parameters(), self.critic1_target.parameters()):
                target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)
            for param, target_param in zip(self.critic2.parameters(), self.critic2_target.parameters()):
                target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)
            for param, target_param in zip(self.actor.parameters(), self.actor_target.parameters()):
                target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)

    def save(self, filename):
        torch.save(self.actor.state_dict(), filename + "_actor.pth")


def train_td3(env, agent, episodes=2000, early_stop_threshold=-150):
    rewards_history = []
    moving_avg = []
    best_avg = -np.inf

    for ep in range(episodes):
        state,_ = env.reset()
        episode_reward = 0
        done = False
        step = 0

        while not done:
            # 线性衰减探索噪声
            if ep < 300:
                agent.exploration_noise = max(0.5 * (1 - ep / 300), 0.1)
            else:
                agent.exploration_noise = 0.1

            action = agent.select_action(state, add_noise=(ep < 100))  # 前100轮强制探索
            next_state, reward, done, _, _ = env.step(action)
            agent.replay_buffer.append((state, action, reward, next_state, done))
            state = next_state
            episode_reward += reward
            agent.train()
            step += 1

        rewards_history.append(episode_reward)
        current_avg = np.mean(rewards_history[-50:])
        moving_avg.append(current_avg)

        if current_avg > best_avg:
            best_avg = current_avg
            agent.save("td3_pendulum_best")

        if (ep + 1) % 50 == 0:
            print(f"Episode: {ep + 1}, Avg Reward: {current_avg:.2f}")

        # 早停机制
        if current_avg >= early_stop_threshold:
            print(f"早停触发，平均奖励达到 {current_avg:.2f}")
            break

    return moving_avg, rewards_history


# 训练并可视化
td3_agent = TD3(state_dim, action_dim, max_action)
moving_avg, rewards_history = train_td3(env, td3_agent, episodes=2000)

# 可视化结果
plt.figure(figsize=(12, 6))
plt.plot(rewards_history, alpha=0.6, label='single round reward')
plt.plot(moving_avg, 'r-', linewidth=2, label='moving average (50 rounds)')
plt.xlabel('episodes')
plt.ylabel('reward')
plt.title('TD3 training performance on Pendulum-v1')
plt.legend()
plt.grid(True)
plt.show()