ddpg的pytorch例程

这是一个使用PyTorch实现DDPG算法的例程：

import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import numpy as np

# 定义Actor网络
class Actor(nn.Module):
    def __init__(self, state_dim, action_dim, max_action):
        super(Actor, self).__init__()
        self.layer1 = nn.Linear(state_dim, 400)
        self.layer2 = nn.Linear(400, 300)
        self.layer3 = nn.Linear(300, action_dim)
        self.max_action = max_action

    def forward(self, x):
        x = F.relu(self.layer1(x))
        x = F.relu(self.layer2(x))
        x = self.max_action * torch.tanh(self.layer3(x))
        return x

# 定义Critic网络
class Critic(nn.Module):
    def __init__(self, state_dim, action_dim):
        super(Critic, self).__init__()
        self.layer1 = nn.Linear(state_dim + action_dim, 400)
        self.layer2 = nn.Linear(400, 300)
        self.layer3 = nn.Linear(300, 1)

    def forward(self, x, u):
        xu = torch.cat([x, u], 1)
        x = F.relu(self.layer1(xu))
        x = F.relu(self.layer2(x))
        x = self.layer3(x)
        return x

# 定义DDPG算法
class DDPG(object):
    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=1e-4)

        self.critic = Critic(state_dim, action_dim).to(device)
        self.critic_target = Critic(state_dim, action_dim).to(device)
        self.critic_target.load_state_dict(self.critic.state_dict())
        self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=1e-3)

        self.max_action = max_action

    def select_action(self, state):
        state = torch.FloatTensor(state.reshape(1, -1)).to(device)
        return self.actor(state).cpu().data.numpy().flatten()

    def train(self, replay_buffer, iterations, batch_size=100, discount=0.99, tau=0.005):
        for it in range(iterations):
            # 从replay buffer中随机采样一个batch的数据
            batch_states, batch_next_states, batch_actions, batch_rewards, batch_dones = replay_buffer.sample(batch_size)
            state = torch.FloatTensor(batch_states).to(device)
            next_state = torch.FloatTensor(batch_next_states).to(device)
            action = torch.FloatTensor(batch_actions).to(device)
            reward = torch.FloatTensor(batch_rewards).to(device)
            done = torch.FloatTensor(batch_dones).to(device)

            # 训练Critic网络
            target_Q = self.critic_target(next_state, self.actor_target(next_state))
            target_Q = reward + ((1 - done) * discount * target_Q).detach()
            current_Q = self.critic(state, action)
            critic_loss = F.mse_loss(current_Q, target_Q)
            self.critic_optimizer.zero_grad()
            critic_loss.backward()
            self.critic_optimizer.step()

            # 训练Actor网络
            actor_loss = -self.critic(state, self.actor(state)).mean()
            self.actor_optimizer.zero_grad()
            actor_loss.backward()
            self.actor_optimizer.step()

            # 更新target网络
            for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()):
                target_param.data.copy_(tau * param.data + (1 - tau) * target_param.data)

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

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

    def load(self, filename):
        self.actor.load_state_dict(torch.load(filename + "_actor"))
        self.critic.load_state_dict(torch.load(filename + "_critic"))
        
# 定义Replay Buffer
class ReplayBuffer(object):
    def __init__(self, max_size=1000000):
        self.storage = []
        self.max_size = max_size
        self.ptr = 0

    def add(self, state, next_state, action, reward, done):
        data = (state, next_state, action, reward, done)
        if len(self.storage) == self.max_size:
            self.storage[int(self.ptr)] = data
            self.ptr = (self.ptr + 1) % self.max_size
        else:
            self.storage.append(data)

    def sample(self, batch_size):
        ind = np.random.randint(0, len(self.storage), size=batch_size)
        batch_states = []
        batch_next_states = []
        batch_actions = []
        batch_rewards = []
        batch_dones = []
        for i in ind:
            state, next_state, action, reward, done = self.storage[i]
            batch_states.append(np.array(state, copy=False))
            batch_next_states.append(np.array(next_state, copy=False))
            batch_actions.append(np.array(action, copy=False))
            batch_rewards.append(np.array(reward, copy=False))
            batch_dones.append(np.array(done, copy=False))
        return np.array(batch_states), np.array(batch_next_states), np.array(batch_actions), np.array(batch_rewards).reshape(-1, 1), np.array(batch_dones).reshape(-1, 1)

# 定义环境
class Environment(object):
    def __init__(self):
        self.state_dim = 3
        self.action_dim = 1
        self.max_action = 2.0

    def step(self, action):
        next_state = np.zeros(3)
        next_state[0] = 0.5
        next_state[1] = 0.5 * np.sin(3 * next_state[0]) + 0.5 * np.sin(2 * next_state[1])
        next_state[2] = 0.5 * np.sin(2 * next_state[0]) + 0.5 * np.sin(3 * next_state[1])
        reward = -np.linalg.norm(next_state - np.array([1, 0, 0])) # 奖励函数为到目标点的欧式距离的相反数
        done = False
        return next_state, reward, done

    def reset(self):
        state = np.zeros(3)
        state[0] = np.random.uniform(low=-0.1, high=0.1)
        state[1] = np.random.uniform(low=-0.1, high=0.1)
        state[2] = np.random.uniform(low=-0.1, high=0.1)
        return state

# 训练DDPG
env = Environment()
replay_buffer = ReplayBuffer()
ddpg = DDPG(env.state_dim, env.action_dim, env.max_action)
for i in range(100000):
    state = env.reset()
    done = False
    while not done:
        action = ddpg.select_action(state)
        next_state, reward, done = env.step(action)
        replay_buffer.add(state, next_state, action, reward, done)
        state = next_state
    if len(replay_buffer.storage) > 1000:
        ddpg.train(replay_buffer, 100)

# 测试DDPG
state = env.reset()
done = False
while not done:
    action = ddpg.select_action(state)
    next_state, reward, done = env.step(action)
    state = next_state
    print(state)
    
# 保存DDPG模型
ddpg.save("ddpg")