|
| 1 | +# %% |
| 2 | +import numpy as np |
| 3 | +import gym |
| 4 | +import torch |
| 5 | +import torch.nn as nn |
| 6 | +import torch.optim as optim |
| 7 | +import random |
| 8 | + |
| 9 | +# 定义Actor网络，适用于连续状态空间 |
| 10 | +class Actor(nn.Module): |
| 11 | + def __init__(self, input_dim, output_dim): |
| 12 | + super(Actor, self).__init__() |
| 13 | + self.net = nn.Sequential( |
| 14 | + nn.Linear(input_dim, 128), |
| 15 | + nn.ReLU(), |
| 16 | + nn.Linear(128, output_dim), |
| 17 | + nn.Tanh() # 使输出在-1到1之间 |
| 18 | + ) |
| 19 | + |
| 20 | + def forward(self, state): |
| 21 | + return self.net(state) |
| 22 | + |
| 23 | +# 定义Critic网络，输出Q值 |
| 24 | +class Critic(nn.Module): |
| 25 | + def __init__(self, input_dim, action_dim): |
| 26 | + super(Critic, self).__init__() |
| 27 | + self.net = nn.Sequential( |
| 28 | + nn.Linear(input_dim + action_dim, 128), |
| 29 | + nn.ReLU(), |
| 30 | + nn.Linear(128, 1) |
| 31 | + ) |
| 32 | + |
| 33 | + def forward(self, state, action): |
| 34 | + return self.net(torch.cat([state, action], dim=1)) |
| 35 | + |
| 36 | +# 经验回放 |
| 37 | +class ReplayBuffer: |
| 38 | + def __init__(self, capacity): |
| 39 | + self.capacity = capacity |
| 40 | + self.buffer = [] |
| 41 | + self.position = 0 |
| 42 | + |
| 43 | + def push(self, state, action, reward, next_state, done): |
| 44 | + if len(self.buffer) < self.capacity: |
| 45 | + self.buffer.append(None) |
| 46 | + self.buffer[self.position] = (state, action, reward, next_state, done) |
| 47 | + self.position = (self.position + 1) % self.capacity |
| 48 | + |
| 49 | + def sample(self, batch_size): |
| 50 | + batch = random.sample(self.buffer, batch_size) |
| 51 | + state, action, reward, next_state, done = map(np.stack, zip(*batch)) |
| 52 | + return state, action, reward, next_state, done |
| 53 | + |
| 54 | + def __len__(self): |
| 55 | + return len(self.buffer) |
| 56 | + |
| 57 | +# %% |
| 58 | +# DDPG更新函数 |
| 59 | +def train(actor, critic, actor_target, critic_target, actor_optimizer, critic_optimizer, replay_buffer, gamma=0.99, tau=0.005): |
| 60 | + state, action, reward, next_state, done = replay_buffer.sample(64) |
| 61 | + |
| 62 | + state = torch.FloatTensor(state) |
| 63 | + action = torch.FloatTensor(action) |
| 64 | + reward = torch.FloatTensor(reward).unsqueeze(1) |
| 65 | + next_state = torch.FloatTensor(next_state) |
| 66 | + done = torch.FloatTensor(done).unsqueeze(1) |
| 67 | + |
| 68 | + # Critic update |
| 69 | + with torch.no_grad(): |
| 70 | + next_action = actor_target(next_state) |
| 71 | + target_q = reward + (1 - done) * gamma * critic_target(next_state, next_action) |
| 72 | + current_q = critic(state, action) |
| 73 | + critic_loss = nn.MSELoss()(current_q, target_q) |
| 74 | + critic_optimizer.zero_grad() |
| 75 | + critic_loss.backward() |
| 76 | + critic_optimizer.step() |
| 77 | + |
| 78 | + # Actor update |
| 79 | + actor_loss = -critic(state, actor(state)).mean() |
| 80 | + actor_optimizer.zero_grad() |
| 81 | + actor_loss.backward() |
| 82 | + actor_optimizer.step() |
| 83 | + |
| 84 | + # Soft update target networks |
| 85 | + for target_param, param in zip(actor_target.parameters(), actor.parameters()): |
| 86 | + target_param.data.copy_(tau * param.data + (1 - tau) * target_param.data) |
| 87 | + for target_param, param in zip(critic_target.parameters(), critic.parameters()): |
| 88 | + target_param.data.copy_(tau * param.data + (1 - tau) * target_param.data) |
| 89 | + |
| 90 | +# %% |
| 91 | +# 主程序 |
| 92 | +if __name__ == "__main__": |
| 93 | + env = gym.make("Pendulum-v1") |
| 94 | + |
| 95 | + actor = Actor(env.observation_space.shape[0], env.action_space.shape[0]) |
| 96 | + critic = Critic(env.observation_space.shape[0], env.action_space.shape[0]) |
| 97 | + |
| 98 | + actor_target = Actor(env.observation_space.shape[0], env.action_space.shape[0]) |
| 99 | + critic_target = Critic(env.observation_space.shape[0], env.action_space.shape[0]) |
| 100 | + |
| 101 | + actor_target.load_state_dict(actor.state_dict()) # 软更新 |
| 102 | + critic_target.load_state_dict(critic.state_dict()) |
| 103 | + |
| 104 | + actor_optimizer = optim.Adam(actor.parameters(), lr=0.001) |
| 105 | + critic_optimizer = optim.Adam(critic.parameters(), lr=0.001) |
| 106 | + |
| 107 | + replay_buffer = ReplayBuffer(1000000) |
| 108 | + |
| 109 | + for episode in range(1000): |
| 110 | + state,_ = env.reset() |
| 111 | + episode_reward = 0 |
| 112 | + |
| 113 | + for step in range(200): # Pendulum-v1默认的最大步数是200 |
| 114 | + action = actor(torch.FloatTensor(state)).detach().numpy() |
| 115 | + next_state, reward, done, _, _ = env.step(action) |
| 116 | + replay_buffer.push(state, action, reward, next_state, done) |
| 117 | + state = next_state |
| 118 | + episode_reward += reward |
| 119 | + |
| 120 | + if len(replay_buffer) > 1000: |
| 121 | + train(actor, critic, actor_target, critic_target, actor_optimizer, critic_optimizer, replay_buffer) |
| 122 | + |
| 123 | + if done: |
| 124 | + break |
| 125 | + |
| 126 | + print(f"Episode {episode}, Reward: {episode_reward}") |
0 commit comments