PyTorch 深度学习实战（15）：Twin Delayed DDPG (TD3) 算法 - IT评测·应用市场-qidao123.com

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()

复制代码