how to

10-actor-critic算法

Nov 11, 2024
notesjulyfun技术学习hrl
3 Minutes
437 Words

see: https://hrl.boyuai.com/chapter/2/actor-critic%E7%AE%97%E6%B3%95

  • 上一章用 $G_t$ 代替 $Q^pi (s, a)$,现在用时序差分残差公式代替之.

    • 因为 $Q = r + gamma V$.
    • 所以训练一个 $V$ 网路就行
  • 原文已经写的很像回忆提纲了

  • 训练一个价值网络:

    • Input : 可微状态 $s$
    • Output : $V(s)$
    • Loss: $$1 / 2 (r + gamma V_omega (s_(t + 1)) - V_omega (s_t))^2$$
      • 其中 $r + gamma V_omega (s_(t + 1))$ 不参与梯度计算. 代码中使用 detach() 直接实现,不用双网络.
      • 和 DQN 一样训练数据来源于采样池.
      • 训练过程和 Actor 的关系?Actor 产生了采样池,Actor 变强后采样分布会变化
        • 采样数据为 $(s_i, a_i, r_i, s_i^prime)$.
        • 注意采样的 $a$ 并不影响 $V$ 梯度下降,乱采样也能训练出正确的 $V$ 网络

先来看 Actor + Critic 包装器的 update

1
class ActorCritic:
2
# self.critic = ValueNet(state_dim, hidden_dim).to(device) # 价值网络
3
...
4
def update(self, transition_dict):
5
states = torch.tensor(transition_dict['states'],
6
dtype=torch.float).to(self.device)
7
actions = torch.tensor(transition_dict['actions']).view(-1, 1).to(
8
self.device)
9
rewards = torch.tensor(transition_dict['rewards'],
10
dtype=torch.float).view(-1, 1).to(self.device)
11
next_states = torch.tensor(transition_dict['next_states'],
12
dtype=torch.float).to(self.device)
13
dones = torch.tensor(transition_dict['dones'],
14
dtype=torch.float).view(-1, 1).to(self.device)
15
35 collapsed lines
16
# 时序差分目标
17
td_target = rewards + self.gamma * self.critic(next_states) * (1 -
18
dones)
19
td_delta = td_target - self.critic(states) # 时序差分误差
20
log_probs = torch.log(self.actor(states).gather(1, actions))
21
actor_loss = torch.mean(-log_probs * td_delta.detach())
22
# 均方误差损失函数,这里直接 detach() 来实现类似 Double DQN 的效果... (不演了是吧)
23
critic_loss = torch.mean(
24
F.mse_loss(self.critic(states), td_target.detach()))
25
self.actor_optimizer.zero_grad()
26
self.critic_optimizer.zero_grad()
27
actor_loss.backward() # 计算策略网络的梯度
28
critic_loss.backward() # 计算价值网络的梯度
29
self.actor_optimizer.step() # 更新策略网络的参数
30
self.critic_optimizer.step() # 更新价值网络的参数
31
32
class PolicyNet(torch.nn.Module):
33
def __init__(self, state_dim, hidden_dim, action_dim):
34
super(PolicyNet, self).__init__()
35
self.fc1 = torch.nn.Linear(state_dim, hidden_dim)
36
self.fc2 = torch.nn.Linear(hidden_dim, action_dim)
37
38
def forward(self, x):
39
x = F.relu(self.fc1(x))
40
return F.softmax(self.fc2(x), dim=1)
41
42
class ValueNet(torch.nn.Module):
43
def __init__(self, state_dim, hidden_dim):
44
super(ValueNet, self).__init__()
45
self.fc1 = torch.nn.Linear(state_dim, hidden_dim)
46
self.fc2 = torch.nn.Linear(hidden_dim, 1)
47
48
def forward(self, x):
49
x = F.relu(self.fc1(x))
50
return self.fc2(x)
  • 效果:抖动比基于蒙特卡洛的 REINFORCE 收敛更快,且非常稳定.
Article title:10-actor-critic算法
Article author:Julyfun
Release time:Nov 11, 2024
Copyright 2025
Sitemap