GCN DDPG 代码

以下是使用 PyTorch 实现的 GCN-DDPG 代码示例：

首先导入所需的库：

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
import random
import math
from collections import namedtuple
from torch_geometric.nn import GCNConv
from torch_geometric.data import Data

定义 Actor 和 Critic 的神经网络模型：

class Actor(nn.Module):
    def __init__(self, num_inputs, num_actions, hidden_size):
        super(Actor, self).__init__()
        self.fc1 = nn.Linear(num_inputs, hidden_size)
        self.fc2 = nn.Linear(hidden_size, hidden_size)
        self.fc3 = nn.Linear(hidden_size, num_actions)
    
    def forward(self, state):
        x = F.relu(self.fc1(state))
        x = F.relu(self.fc2(x))
        x = torch.tanh(self.fc3(x))
        return x

class Critic(nn.Module):
    def __init__(self, num_inputs, num_actions, hidden_size):
        super(Critic, self).__init__()
        self.fc1 = nn.Linear(num_inputs + num_actions, hidden_size)
        self.fc2 = nn.Linear(hidden_size, hidden_size)
        self.fc3 = nn.Linear(hidden_size, 1)
    
    def forward(self, state, action):
        x = torch.cat([state, action], 1)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

定义 GCN-DDPG 的智能体：

class Agent(object):
    def __init__(self, env, num_hidden=64, gamma=0.99, tau=0.001, lr=0.001, buffer_size=10000, batch_size=64):
        self.env = env
        self.observation_space = env.observation_space
        self.action_space = env.action_space
        self.num_inputs = self.observation_space.shape[0]
        self.num_actions = self.action_space.shape[0]
        self.hidden_size = num_hidden
        self.gamma = gamma
        self.tau = tau
        self.lr = lr
        self.buffer_size = buffer_size
        self.batch_size = batch_size
        self.memory = []
        self.timestep = 0
        self.actor = Actor(self.num_inputs, self.num_actions, self.hidden_size).to(device)
        self.actor_target = Actor(self.num_inputs, self.num_actions, self.hidden_size).to(device)
        self.critic = Critic(self.num_inputs, self.num_actions, self.hidden_size).to(device)
        self.critic_target = Critic(self.num_inputs, self.num_actions, self.hidden_size).to(device)
        self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=self.lr)
        self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=self.lr)
    
    def select_action(self, state):
        state = torch.from_numpy(state).float().to(device)
        with torch.no_grad():
            action = self.actor(state).cpu().data.numpy()
        return action
    
    def store_transition(self, state, action, reward, next_state, done):
        self.memory.append((state, action, reward, next_state, done))
        if len(self.memory) > self.buffer_size:
            self.memory.pop(0)
    
    def update(self):
        if len(self.memory) < self.batch_size:
            return
        batch = random.sample(self.memory, self.batch_size)
        state_batch = torch.FloatTensor([transition[0] for transition in batch]).to(device)
        action_batch = torch.FloatTensor([transition[1] for transition in batch]).to(device)
        reward_batch = torch.FloatTensor([transition[2] for transition in batch]).to(device)
        next_state_batch = torch.FloatTensor([transition[3] for transition in batch]).to(device)
        done_batch = torch.FloatTensor([transition[4] for transition in batch]).to(device)
        
        # Update Critic
        next_action_batch = self.actor_target(next_state_batch)
        q_next = self.critic_target(next_state_batch, next_action_batch)
        q_target = reward_batch + self.gamma * q_next * (1 - done_batch)
        q_current = self.critic(state_batch, action_batch)
        critic_loss = F.mse_loss(q_current, q_target)
        self.critic_optimizer.zero_grad()
        critic_loss.backward()
        self.critic_optimizer.step()
        
        # Update Actor
        action_pred = self.actor(state_batch)
        actor_loss = -self.critic(state_batch, action_pred).mean()
        self.actor_optimizer.zero_grad()
        actor_loss.backward()
        self.actor_optimizer.step()
        
        # Soft Update of Target Networks
        for target_param, param in zip(self.actor_target.parameters(), self.actor.parameters()):
            target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)
        for target_param, param in zip(self.critic_target.parameters(), self.critic.parameters()):
            target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)

定义 GCN-DDPG 的环境：

class GraphEnvironment(object):
    def __init__(self, num_nodes=10, num_edges=20, obs_dim=2):
        self.num_nodes = num_nodes
        self.num_edges = num_edges
        self.obs_dim = obs_dim
        self.adj_matrix = self._generate_adj_matrix()
        self.observation_space = np.zeros((self.num_nodes, self.obs_dim))
        self.action_space = np.zeros((self.num_nodes, self.obs_dim))
    
    def _generate_adj_matrix(self):
        edge_list = []
        while len(edge_list) < self.num_edges:
            src = random.randint(0, self.num_nodes - 1)
            dst = random.randint(0, self.num_nodes - 1)
            if src != dst and (src, dst) not in edge_list and (dst, src) not in edge_list:
                edge_list.append((src, dst))
        adj_matrix = np.zeros((self.num_nodes, self.num_nodes))
        for src, dst in edge_list:
            adj_matrix[src][dst] = 1
            adj_matrix[dst][src] = 1
        return adj_matrix
    
    def reset(self):
        self.observation_space = np.random.rand(self.num_nodes, self.obs_dim)
        self.action_space = np.zeros((self.num_nodes, self.obs_dim))
        return self.observation_space
    
    def step(self, action):
        self.action_space = action
        reward = 0
        for i in range(self.num_nodes):
            neighbors = np.where(self.adj_matrix[i] == 1)[0]
            for j in neighbors:
                reward += np.linalg.norm(self.observation_space[i] - self.observation_space[j]) - np.linalg.norm(self.action_space[i] - self.action_space[j])
        self.observation_space += self.action_space
        return self.observation_space, reward, False, {}

定义训练过程：

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

env = GraphEnvironment()
agent = Agent(env, num_hidden=64, gamma=0.99, tau=0.001, lr=0.001, buffer_size=10000, batch_size=64)

num_episodes = 1000
for i_episode in range(num_episodes):
    state = env.reset()
    episode_reward = 0
    for t in range(100):
        action = agent.select_action(state)
        next_state, reward, done, _ = env.step(action)
        agent.store_transition(state, action, reward, next_state, done)
        agent.update()
        state = next_state
        episode_reward += reward
        if done:
            break
    print("Episode: {}, Reward: {}, Timestep: {}".format(i_episode + 1, episode_reward, agent.timestep))

这是一个简单的 GCN-DDPG 实现，仅供参考。具体实现可能需要根据不同的问题和数据进行调整。