深度强化学习-TD3算法

论文地址：https://arxiv.org/pdf/1802.09477.pdf

        TD3（Twin Delayed Deep Deterministic policy gradient algorithm）算法适合于高维连续动作空间，是DDPG算法的优化版本，为了优化DDPG在训练过程中Q值估计过高的问题。

相较DDPG的改进：

1、运用两个Critic网络。运用两个网络对动作价值函数进行估计。在练习的时分挑选最小的Q值作为估值（为了防止误差累积过高）。

2、运用延迟学习。Critic网络更新的频率要比Actor网络更新的频率要大（类似GAN的思想，先训练好Critic才能更好的对actor指指点点）。

3、运用梯度截取。将Actor的参数更新的梯度截取到某个范围内。

4、加入训练噪声。更新Critic网络时候加入随机噪声，以达到对Critic网络波动的稳定性。

算法流程：

        算法的伪代码 

代码实现： 

        actor：

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

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

        self.maxaction = maxaction

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

         critic：

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

        # Q1 architecture
        self.l1 = nn.Linear(state_dim + action_dim, net_width)
        self.l2 = nn.Linear(net_width, net_width)
        self.l3 = nn.Linear(net_width, 1)

        # Q2 architecture
        self.l4 = nn.Linear(state_dim + action_dim, net_width)
        self.l5 = nn.Linear(net_width, net_width)
        self.l6 = nn.Linear(net_width, 1)

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

        q1 = F.relu(self.l1(sa))
        q1 = F.relu(self.l2(q1))
        q1 = self.l3(q1)

        q2 = F.relu(self.l4(sa))
        q2 = F.relu(self.l5(q2))
        q2 = self.l6(q2)
        return q1, q2

    def Q1(self, state, action):
        sa = torch.cat([state, action], 1)

        q1 = F.relu(self.l1(sa))
        q1 = F.relu(self.l2(q1))
        q1 = self.l3(q1)
        return q1

         TD3的整体实现：

class TD3(object):
    def __init__(
            self,
            env_with_Dead,
            state_dim,
            action_dim,
            max_action,
            gamma=0.99,
            net_width=128,
            a_lr=1e-4,
            c_lr=1e-4,
            Q_batchsize=256
    ):

        self.actor = Actor(state_dim, action_dim, net_width, max_action).to(device)
        self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=a_lr)
        self.actor_target = copy.deepcopy(self.actor)

        self.q_critic = Q_Critic(state_dim, action_dim, net_width).to(device)
        self.q_critic_optimizer = torch.optim.Adam(self.q_critic.parameters(), lr=c_lr)
        self.q_critic_target = copy.deepcopy(self.q_critic)

        self.env_with_Dead = env_with_Dead
        self.action_dim = action_dim
        self.max_action = max_action
        self.gamma = gamma
        self.policy_noise = 0.2 * max_action
        self.noise_clip = 0.5 * max_action
        self.tau = 0.005
        self.Q_batchsize = Q_batchsize
        self.delay_counter = -1
        self.delay_freq = 1

    def select_action(self, state):  # only used when interact with the env
        with torch.no_grad():
            state = torch.FloatTensor(state.reshape(1, -1)).to(device)
            a = self.actor(state)
        return a.cpu().numpy().flatten()

    def train(self, replay_buffer):
        self.delay_counter += 1
        with torch.no_grad():
            s, a, r, s_prime, dead_mask = replay_buffer.sample(self.Q_batchsize)
            noise = (torch.randn_like(a) * self.policy_noise).clamp(-self.noise_clip, self.noise_clip)
            smoothed_target_a = (
                    self.actor_target(s_prime) + noise  # Noisy on target action
            ).clamp(-self.max_action, self.max_action)

        # Compute the target Q value
        target_Q1, target_Q2 = self.q_critic_target(s_prime, smoothed_target_a)
        target_Q = torch.min(target_Q1, target_Q2)
        '''DEAD OR NOT'''
        if self.env_with_Dead:
            target_Q = r + (1 - dead_mask) * self.gamma * target_Q  # env with dead
        else:
            target_Q = r + self.gamma * target_Q  # env without dead

        # Get current Q estimates
        current_Q1, current_Q2 = self.q_critic(s, a)

        # Compute critic loss
        q_loss = F.mse_loss(current_Q1, target_Q) + F.mse_loss(current_Q2, target_Q)

        # Optimize the q_critic
        self.q_critic_optimizer.zero_grad()
        q_loss.backward()
        self.q_critic_optimizer.step()

        if self.delay_counter == self.delay_freq:
            # Update Actor
            a_loss = -self.q_critic.Q1(s, self.actor(s)).mean()
            self.actor_optimizer.zero_grad()
            a_loss.backward()
            self.actor_optimizer.step()

            # Update the frozen target models
            for param, target_param in zip(self.q_critic.parameters(), self.q_critic_target.parameters()):
                target_param.data.copy_(self.tau * param.data + (1 - self.tau) * target_param.data)

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

            self.delay_counter = -1

    def save(self, episode):
        torch.save(self.actor.state_dict(), "ppo_actor{}.pth".format(episode))
        torch.save(self.q_critic.state_dict(), "ppo_q_critic{}.pth".format(episode))

    def load(self, episode):

        self.actor.load_state_dict(torch.load("ppo_actor{}.pth".format(episode)))
        self.q_critic.load_state_dict(torch.load("ppo_q_critic{}.pth".format(episode)))

 网络结构图：

         其中actor和target部分的网络参数会延迟更新，也就是说critic1和critic2参数在不断更新，训练好critic之后才能知道actor做出理想的动作。