DDPG强化学习的PyTorch代码实现和逐步讲解

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

作者：Renu Khandelwal