DDPG强化学习的PyTorch代码实现和逐步讲解

来源：Deephub Imba
本文约4300字，建议阅读10分钟
本文将使用pytorch对其进行完整的实现和讲解。

深度确定性策略梯度(Deep Deterministic Policy Gradient, DDPG)是受Deep Q-Network启发的无模型、非策略深度强化算法，是基于使用策略梯度的Actor-Critic，本文将使用pytorch对其进行完整的实现和讲解。

DDPG的关键组成部分是

• Replay Buffer

• Actor-Critic neural network

• Exploration Noise

• Target network

• Soft Target Updates for Target Network

下面我们一个一个来逐步实现：

Replay Buffer

DDPG使用Replay Buffer存储通过探索环境采样的过程和奖励(Sₜ，aₜ，Rₜ，Sₜ+₁)。Replay Buffer在帮助代理加速学习以及DDPG的稳定性方面起着至关重要的作用：

• 最小化样本之间的相关性：将过去的经验存储在 Replay Buffer 中，从而允许代理从各种经验中学习。

• 启用离线策略学习：允许代理从重播缓冲区采样转换，而不是从当前策略采样转换。

• 高效采样：将过去的经验存储在缓冲区中，允许代理多次从不同的经验中学习。

class Replay_buffer():
     '''
    Code based on:
    https://github.com/openai/baselines/blob/master/baselines/deepq/replay_buffer.py
    Expects tuples of (state, next_state, action, reward, done)
    '''
     def __init__(self, max_size=capacity):
         """Create Replay buffer.
        Parameters
        ----------
        size: int
            Max number of transitions to store in the buffer. When the buffer
            overflows the old memories are dropped.
        """
         self.storage = []
         self.max_size = max_size
         self.ptr = 0


     def push(self, data):
         if len(self.storage) == self.max_size:
             self.storage[int(self.ptr)] = data
             self.ptr = (self.ptr + 1) % self.max_size
         else:
             self.storage.append(data)


     def sample(self, batch_size):
         """Sample a batch of experiences.
        Parameters
        ----------
        batch_size: int
            How many transitions to sample.
        Returns
        -------
        state: np.array
            batch of state or observations
        action: np.array
            batch of actions executed given a state
        reward: np.array
            rewards received as results of executing action
        next_state: np.array
            next state next state or observations seen after executing action
        done: np.array
            done[i] = 1 if executing ation[i] resulted in
            the end of an episode and 0 otherwise.
        """
         ind = np.random.randint(0, len(self.storage), size=batch_size)
         state, next_state, action, reward, done = [], [], [], [], []


         for i in ind:
             st, n_st, act, rew, dn = self.storage[i]
             state.append(np.array(st, copy=False))
             next_state.append(np.array(n_st, copy=False))
             action.append(np.array(act, copy=False))
             reward.append(np.array(rew, copy=False))
             done.append(np.array(dn, copy=False))


         return np.array(state), np.array(next_state), np.array(action), np.array(reward).reshape(-1, 1), np.array(done).reshape(-1, 1)

Actor-Critic Neural Network

这是Actor-Critic 强化学习算法的 PyTorch 实现。该代码定义了两个神经网络模型，一个 Actor 和一个 Critic。

Actor 模型的输入：环境状态；Actor 模型的输出：具有连续值的动作。

Critic 模型的输入：环境状态和动作；Critic 模型的输出：Q 值，即当前状态-动作对的预期总奖励。

class Actor(nn.Module):
     """
    The Actor model takes in a state observation as input and
    outputs an action, which is a continuous value.


    It consists of four fully connected linear layers with ReLU activation functions and
    a final output layer selects one single optimized action for the state
    """
     def __init__(self, n_states, action_dim, hidden1):
         super(Actor, self).__init__()
         self.net = nn.Sequential(
             nn.Linear(n_states, hidden1),
             nn.ReLU(),
             nn.Linear(hidden1, hidden1),
             nn.ReLU(),
             nn.Linear(hidden1, hidden1),
             nn.ReLU(),
             nn.Linear(hidden1, 1)
        )


     def forward(self, state):
         return self.net(state)


 class Critic(nn.Module):
     """
    The Critic model takes in both a state observation and an action as input and
    outputs a Q-value, which estimates the expected total reward for the current state-action pair.


    It consists of four linear layers with ReLU activation functions,
    State and action inputs are concatenated before being fed into the first linear layer.


    The output layer has a single output, representing the Q-value
    """
     def __init__(self, n_states, action_dim, hidden2):
         super(Critic, self).__init__()
         self.net = nn.Sequential(
             nn.Linear(n_states + action_dim, hidden2),
             nn.ReLU(),
             nn.Linear(hidden2, hidden2),
             nn.ReLU(),
             nn.Linear(hidden2, hidden2),
             nn.ReLU(),
             nn.Linear(hidden2, action_dim)
        )


     def forward(self, state, action):
         return self.net(torch.cat((state, action), 1))

Exploration Noise

向 Actor 选择的动作添加噪声是 DDPG 中用来鼓励探索和改进学习过程的一种技术。

可以使用高斯噪声或 Ornstein-Uhlenbeck 噪声。 高斯噪声简单且易于实现，Ornstein-Uhlenbeck 噪声会生成时间相关的噪声，可以帮助代理更有效地探索动作空间。但是与高斯噪声方法相比，Ornstein-Uhlenbeck 噪声波动更平滑且随机性更低。

import numpy as np
 import random
 import copy


 class OU_Noise(object):
     """Ornstein-Uhlenbeck process.
    code from :
    https://math.stackexchange.com/questions/1287634/implementing-ornstein-uhlenbeck-in-matlab
    The OU_Noise class has four attributes


        size: the size of the noise vector to be generated
        mu: the mean of the noise, set to 0 by default
        theta: the rate of mean reversion, controlling how quickly the noise returns to the mean
        sigma: the volatility of the noise, controlling the magnitude of fluctuations
    """
     def __init__(self, size, seed, mu=0., theta=0.15, sigma=0.2):
         self.mu = mu * np.ones(size)
         self.theta = theta
         self.sigma = sigma
         self.seed = random.seed(seed)
         self.reset()


     def reset(self):
         """Reset the internal state (= noise) to mean (mu)."""
         self.state = copy.copy(self.mu)


     def sample(self):
         """Update internal state and return it as a noise sample.
        This method uses the current state of the noise and generates the next sample
        """
         dx = self.theta * (self.mu - self.state) + self.sigma * np.array([np.random.normal() for _ in range(len(self.state))])
         self.state += dx
         return self.state

要在DDPG中使用高斯噪声，可以直接将高斯噪声添加到代理的动作选择过程中。

DDPG

DDPG (Deep Deterministic Policy Gradient)采用两组Actor-Critic神经网络进行函数逼近。在DDPG中，目标网络是Actor-Critic ，它目标网络具有与Actor-Critic网络相同的结构和参数化。

在训练期时，代理使用其 Actor-Critic 网络与环境交互，并将经验元组（Sₜ、Aₜ、Rₜ、Sₜ+₁）存储在Replay Buffer中。 然后代理从 Replay Buffer 中采样并使用数据更新 Actor-Critic 网络。 DDPG 算法不是通过直接从 Actor-Critic 网络复制来更新目标网络权重，而是通过称为软目标更新的过程缓慢更新目标网络权重。

软目标的更新是从Actor-Critic网络传输到目标网络的称为目标更新率(τ)的权重的一小部分。

软目标的更新公式如下:

通过使用软目标技术，可以大大提高学习的稳定性。

#Set Hyperparameters
 # Hyperparameters adapted for performance from
 capacity=1000000
 batch_size=64
 update_iteration=200
 tau=0.001 # tau for soft updating
 gamma=0.99 # discount factor
 directory = './'
 hidden1=20 # hidden layer for actor
 hidden2=64. #hiiden laye for critic


 class DDPG(object):
     def __init__(self, state_dim, action_dim):
         """
        Initializes the DDPG agent.
        Takes three arguments:
                state_dim which is the dimensionality of the state space,
                action_dim which is the dimensionality of the action space, and
                max_action which is the maximum value an action can take.


        Creates a replay buffer, an actor-critic networks and their corresponding target networks.
        It also initializes the optimizer for both actor and critic networks alog with
        counters to track the number of training iterations.
        """
         self.replay_buffer = Replay_buffer()


         self.actor = Actor(state_dim, action_dim, hidden1).to(device)
         self.actor_target = Actor(state_dim, action_dim,  hidden1).to(device)
         self.actor_target.load_state_dict(self.actor.state_dict())
         self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=3e-3)


         self.critic = Critic(state_dim, action_dim,  hidden2).to(device)
         self.critic_target = Critic(state_dim, action_dim,  hidden2).to(device)
         self.critic_target.load_state_dict(self.critic.state_dict())
         self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=2e-2)
         # learning rate






         self.num_critic_update_iteration = 0
         self.num_actor_update_iteration = 0
         self.num_training = 0


     def select_action(self, state):
         """
        takes the current state as input and returns an action to take in that state.
        It uses the actor network to map the state to an action.
        """
         state = torch.FloatTensor(state.reshape(1, -1)).to(device)
         return self.actor(state).cpu().data.numpy().flatten()




     def update(self):
         """
        updates the actor and critic networks using a batch of samples from the replay buffer.
        For each sample in the batch, it computes the target Q value using the target critic network and the target actor network.
        It then computes the current Q value
        using the critic network and the action taken by the actor network.


        It computes the critic loss as the mean squared error between the target Q value and the current Q value, and
        updates the critic network using gradient descent.


        It then computes the actor loss as the negative mean Q value using the critic network and the actor network, and
        updates the actor network using gradient ascent.


        Finally, it updates the target networks using
        soft updates, where a small fraction of the actor and critic network weights are transferred to their target counterparts.
        This process is repeated for a fixed number of iterations.
        """


         for it in range(update_iteration):
             # For each Sample in replay buffer batch
             state, next_state, action, reward, done = self.replay_buffer.sample(batch_size)
             state = torch.FloatTensor(state).to(device)
             action = torch.FloatTensor(action).to(device)
             next_state = torch.FloatTensor(next_state).to(device)
             done = torch.FloatTensor(1-done).to(device)
             reward = torch.FloatTensor(reward).to(device)


             # Compute the target Q value
             target_Q = self.critic_target(next_state, self.actor_target(next_state))
             target_Q = reward + (done * gamma * target_Q).detach()


             # Get current Q estimate
             current_Q = self.critic(state, action)


             # Compute critic loss
             critic_loss = F.mse_loss(current_Q, target_Q)


             # Optimize the critic
             self.critic_optimizer.zero_grad()
             critic_loss.backward()
             self.critic_optimizer.step()


             # Compute actor loss as the negative mean Q value using the critic network and the actor network
             actor_loss = -self.critic(state, self.actor(state)).mean()


             # Optimize the actor
             self.actor_optimizer.zero_grad()
             actor_loss.backward()
             self.actor_optimizer.step()




             """
            Update the frozen target models using
            soft updates, where
            tau,a small fraction of the actor and critic network weights are transferred to their target counterparts.
            """
             for param, target_param in zip(self.critic.parameters(), self.critic_target.parameters()):
                 target_param.data.copy_(tau * param.data + (1 - tau) * target_param.data)


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




             self.num_actor_update_iteration += 1
             self.num_critic_update_iteration += 1
     def save(self):
         """
        Saves the state dictionaries of the actor and critic networks to files
        """
         torch.save(self.actor.state_dict(), directory + 'actor.pth')
         torch.save(self.critic.state_dict(), directory + 'critic.pth')


     def load(self):
         """
        Loads the state dictionaries of the actor and critic networks to files
        """
         self.actor.load_state_dict(torch.load(directory + 'actor.pth'))
         self.critic.load_state_dict(torch.load(directory + 'critic.pth'))

训练DDPG

这里我们使用 OpenAI Gym 的“MountainCarContinuous-v0”来训练我们的DDPG RL 模型，这里的环境提供连续的行动和观察空间，目标是尽快让小车到达山顶。

下面定义算法的各种参数，例如最大训练次数、探索噪声和记录间隔等等。 使用固定的随机种子可以使得过程能够回溯。

import gym


 # create the environment
 env_name='MountainCarContinuous-v0'
 env = gym.make(env_name)
 device = 'cuda' if torch.cuda.is_available() else 'cpu'


 # Define different parameters for training the agent
 max_episode=100
 max_time_steps=5000
 ep_r = 0
 total_step = 0
 score_hist=[]
 # for rensering the environmnet
 render=True
 render_interval=10
 # for reproducibility
 env.seed(0)
 torch.manual_seed(0)
 np.random.seed(0)
 #Environment action ans states
 state_dim = env.observation_space.shape[0]
 action_dim = env.action_space.shape[0]
 max_action = float(env.action_space.high[0])
 min_Val = torch.tensor(1e-7).float().to(device)


 # Exploration Noise
 exploration_noise=0.1
 exploration_noise=0.1 * max_action

创建DDPG代理类的实例，以训练代理达到指定的次数。在每轮结束时调用代理的update()方法来更新参数，并且在每十轮之后使用save()方法将代理的参数保存到一个文件中。

# Create a DDPG instance
 agent = DDPG(state_dim, action_dim)


 # Train the agent for max_episodes
 for i in range(max_episode):
     total_reward = 0
     step =0
     state = env.reset()
     for  t in range(max_time_steps):
         action = agent.select_action(state)
         # Add Gaussian noise to actions for exploration
         action = (action + np.random.normal(0, 1, size=action_dim)).clip(-max_action, max_action)
         #action += ou_noise.sample()
         next_state, reward, done, info = env.step(action)
         total_reward += reward
         if render and i >= render_interval : env.render()
         agent.replay_buffer.push((state, next_state, action, reward, np.float(done)))
         state = next_state
         if done:
             break
         step += 1


     score_hist.append(total_reward)
     total_step += step+1
     print("Episode: \t{} Total Reward: \t{:0.2f}".format( i, total_reward))
     agent.update()
     if i % 10 == 0:
         agent.save()
 env.close()

测试DDPG

test_iteration=100


 for i in range(test_iteration):
     state = env.reset()
     for t in count():
         action = agent.select_action(state)
         next_state, reward, done, info = env.step(np.float32(action))
         ep_r += reward
         print(reward)
         env.render()
         if done:
             print("reward{}".format(reward))
             print("Episode \t{}, the episode reward is \t{:0.2f}".format(i, ep_r))
             ep_r = 0
             env.render()
             break
         state = next_state

我们使用下面的参数让模型收敛：

• 从标准正态分布中采样噪声，而不是随机采样。

• 将polyak常数(tau)从0.99更改为0.001

• 修改Critic 网络的隐藏层大小为[64,64]。在Critic 网络的第二层之后删除了ReLU激活。改成(Linear, ReLU, Linear, Linear)。

• 最大缓冲区大小更改为1000000

• 将batch_size的大小从128更改为64

训练了75轮之后的效果如下：

总结

DDPG算法是一种受deep Q-Network (DQN)算法启发的无模型off-policy Actor-Critic算法。它结合了策略梯度方法和Q-learning的优点来学习连续动作空间的确定性策略。

与DQN类似，它使用重播缓冲区存储过去的经验和目标网络，用于训练网络，从而提高了训练过程的稳定性。

DDPG算法需要仔细的超参数调优以获得最佳性能。超参数包括学习率、批大小、目标网络更新速率和探测噪声参数。超参数的微小变化会对算法的性能产生重大影响。

上面的参数来自：

https://ai.stackexchange.com/questions/22945/ddpg-doesnt-converge-for-mountaincarcontinuous-v0-gym-environment

本文的完整代码：

https://github.com/arshren/Reinforcement_Learning/blob/main/DDPG-MountainCar.ipynb

编辑：王菁

校对：林亦霖