如何正确优化演员和评论家之间的共享网络？

Question

我正在构建一个演员批评强化学习算法来解决环境问题。我想使用单个编码器来查找我的环境的表示。

当我与演员和评论家共享编码器时，我的网络没有学习任何东西：

class Encoder(nn.Module):
  def __init__(self, state_dim):
    super(Encoder, self).__init__()

    self.l1 = nn.Linear(state_dim, 512)

  def forward(self, state):
    a = F.relu(self.l1(state))
    return a

class Actor(nn.Module):
  def __init__(self, state_dim, action_dim, max_action):
    super(Actor, self).__init__()

    self.l1 = nn.Linear(state_dim, 128)
    self.l3 = nn.Linear(128, action_dim)

    self.max_action = max_action

  def forward(self, state):
    a = F.relu(self.l1(state))
    # a = F.relu(self.l2(a))
    a = torch.tanh(self.l3(a)) * self.max_action
    return a

class Critic(nn.Module):
  def __init__(self, state_dim, action_dim):
    super(Critic, self).__init__()

    self.l1 = nn.Linear(state_dim + action_dim, 128)
    self.l3 = nn.Linear(128, 1)

  def forward(self, state, action):
    state_action = torch.cat([state, action], 1)

    q = F.relu(self.l1(state_action))
    # q = F.relu(self.l2(q))
    q = self.l3(q)
    return q

但是，当我为演员使用不同的编码器并为评论家使用不同的编码器时，它会正确学习。

class Actor(nn.Module):
def __init__(self, state_dim, action_dim, max_action):
    super(Actor, self).__init__()

    self.l1 = nn.Linear(state_dim, 400)
    self.l2 = nn.Linear(400, 300)
    self.l3 = nn.Linear(300, action_dim)

    self.max_action = max_action

def forward(self, state):
    a = F.relu(self.l1(state))
    a = F.relu(self.l2(a))
    a = torch.tanh(self.l3(a)) * self.max_action
    return a

class Critic(nn.Module):
  def __init__(self, state_dim, action_dim):
    super(Critic, self).__init__()

    self.l1 = nn.Linear(state_dim + action_dim, 400)
    self.l2 = nn.Linear(400, 300)
    self.l3 = nn.Linear(300, 1)

  def forward(self, state, action):
    state_action = torch.cat([state, action], 1)

    q = F.relu(self.l1(state_action))
    q = F.relu(self.l2(q))
    q = self.l3(q)
    return q

我很确定它是因为优化器。 在共享编码器代码中，我将其定义为傻瓜：

self.actor_optimizer = optim.Adam(list(self.actor.parameters())+
                                      list(self.encoder.parameters()))
self.critic_optimizer = optim.Adam(list(self.critic.parameters()))
                                         +list(self.encoder.parameters()))

在单独的编码器中，它只是：

self.actor_optimizer = optim.Adam((self.actor.parameters()))
self.critic_optimizer = optim.Adam((self.critic.parameters()))

两个优化器必须是因为actorcritic算法。

如何结合两个优化器来正确优化编码器？

Answer 1

我不确定你是如何共享编码器的。

但是，我建议您创建一个编码器实例并将其传递给演员和评论家

encoder_net = Encoder(state_dim)
actor = Actor(encoder_net, state_dim, action_dim, max_action)
critic = Critic(encoder_net, state_dim)

在前向传递期间，首先将状态批次先通过编码器，然后再通过网络的其余部分，例如：

class Encoder(nn.Module):
    def __init__(self, state_dim):
        super(Encoder, self).__init__()

        self.l1 = nn.Linear(state_dim, 512)

    def forward(self, state):
        a = F.relu(self.l1(state))
        return a

class Actor(nn.Module):
    def __init__(self, encoder, state_dim, action_dim, max_action):
        super(Actor, self).__init__()
        self.encoder = encoder

        self.l1 = nn.Linear(512, 128)
        self.l3 = nn.Linear(128, action_dim)

        self.max_action = max_action

    def forward(self, state):
        state = self.encoder(state)
        a = F.relu(self.l1(state))
        # a = F.relu(self.l2(a))
        a = torch.tanh(self.l3(a)) * self.max_action
        return a

class Critic(nn.Module):
    def __init__(self, encoder, state_dim):
        super(Critic, self).__init__()
        self.encoder = encoder

        self.l1 = nn.Linear(512, 128)
        self.l3 = nn.Linear(128, 1)

    def forward(self, state):
        state = self.encoder(state)

        q = F.relu(self.l1(state))
        # q = F.relu(self.l2(q))
        q = self.l3(q)
        return q

注意：critic 网络现在是状态值函数 V(s) 的函数逼近器，而不是状态-动作值函数 Q(s,a)。

使用此实现，您可以在不将编码器参数传递给优化器的情况下执行优化，如下所示：

self.actor_optimizer = optim.Adam((self.actor.parameters()))
self.critic_optimizer = optim.Adam((self.critic.parameters()))

因为编码器参数现在在两个网络之间共享。

希望这会有所帮助！ :)