PPO算法车杆环境倒立摆环境

最新推荐文章于 2024-06-18 13:11:16 发布

小点点~~

最新推荐文章于 2024-06-18 13:11:16 发布

阅读量877

点赞数 20

文章标签：算法人工智能 pytorch

本文链接：https://blog.csdn.net/oceancoco/article/details/134717929

版权

第十二章 PPO算法

12.1 简介

上一章介绍的TRPO算法在很多场景上的应用都很成功，但是它的计算过程非常复杂，每一步更新的运算量非常大。于是，TRPO算法的改进版–PPO算法在2017年被提出，PPO基于TRPO的思想，但是其算法实现更加简单。并且大量试验结果表明，与TRPO算法相比，PPO能学习得一样好（甚至更快），这使得PPO成为非常流行的强化学习算法。如果我们想要尝试在一个新的环境中使用强化学习算法，那么PPO就属于可以首先尝试的算法。

TRPO算法的优化目标：

$\underset{\theta}\max \mathbb E_{s~\nu ^{\pi _{\theta_k }}}\mathbb E_{a~\pi _{\theta_k }\left( \cdot |s \right)}\left[\frac{\pi_{\theta_k}(a|s)}{\pi_\theta(a|s)} A^{\pi _{\theta_k}}\left( s,a \right) \right]$

$\mathbb E_{s~\nu ^{\pi _{\theta _k}}}\left[ D_{KL}\left( \pi _{\theta _k}\left( \cdot |s \right) ,\pi _{\theta }\left( \cdot |s \right) \right) \right] \le \delta$

TRPO使用泰勒展开近似、共轭梯度、线性搜索等方法直接求解。PPO的优化目标和TRPO算法的相同，但是PPO算法用了一些相对简单的方法来求解。具体来说PPO算法有两种形式，一是PPO-惩罚，二是PPO-截断，以下是对这两种形式的介绍。

12.2 PPO-惩罚

PPO-惩罚（PPO-Penalty）用拉格朗日乘数法直接将KL散度的限制放进了目标函数中，这就变成了一个无约束的优化问题，在迭代过程中不断更新KL散度前的系数。即：

$arg\underset{\theta}\max \mathbb E_{s~\nu ^{\pi _{\theta_k }}}\mathbb E_{a~\pi _{\theta_k }\left( \cdot |s \right)}\left[\frac{\pi_{\theta_k}(a|s)}{\pi_\theta(a|s)} A^{\pi _{\theta_k}}\left( s,a \right) -\beta D_{KL}[\pi_{\theta_k}(\cdot|s),\pi_\theta(\cdot|s)]\right]$

令 $d_k=D^{\nu^{\pi_{\theta_k}}}_{KL}(\pi_{\theta_k},\pi_\theta)$ ， $\beta$ 的更新规则如下：

如果 $d_k< \delta /1.5$ ，那么 $\beta_{k+1}=\beta _k/2$
如果 $d_k>\delta ×1.5$ ，那么 $\beta_{k+1}=\beta _k×2$
否则 $\beta_{k+1}=\beta_k$

其中， $\delta$ 是事先设定的一个超参数，用于限制学习策略和之前一轮策略的差距。

12.3 PPO-截断

PPO的另一种形式PPO-截断（PPO-Clip）更加直接，它在目标函数中进行限制，以保证新的参数和旧的参数的差距不会太大，即：

$arg\underset{\theta}\max \mathbb E_{s~\nu ^{\pi _{\theta_k }}}\mathbb E_{a~\pi _{\theta_k }\left( \cdot |s \right)}\left[min(\frac{\pi_{\theta_k}(a|s)}{\pi_\theta(a|s)} A^{\pi _{\theta_k}}\left( s,a \right),clip(\frac{\pi_{\theta_k}(a|s)}{\pi_\theta(a|s)},1-\epsilon,1+\epsilon)A^{\pi _{\theta_k}}\right]$

其中 $c l i p (x, l, r) := ma x (min (x, r), l)$ ，即把 $x$ 限制在 $[l, r]$ 内。上式中 $\epsilon$ 是一个超参数，表示进行截断（clip）的范围。

如果 $A^{\pi _{\theta_k}}(s,a)>0$ 说明这个动作的价值高于平均，最大化会增大 $\frac{\pi_\theta(a|s)}{\pi_{\theta_k}(a|s)}$ ，但不会让其超过 $1+\epsilon$ 。反之，如果 $A^{\pi _{\theta_k}}(s,a)<0$ ，最大化这个式子会减小 $\frac{\pi_\theta(a|s)}{\pi_{\theta_k}(a|s)}$ ，但是不会让其超过 $1-\epsilon$ 。如图12-1所示：

在这里插入图片描述

12.4 PPO代码实践

与TRPO相同，我们仍然在车杆和倒立摆两个环境中测试PPO算法。大量实验表明，PPO-截断总是比PPO-惩罚表现得更好。因此下面我们专注于PPO-截断的代码实现。

12.4.1 车杆环境

import gym
import torch
import torch.nn.functional as F
import numpy as np
import matplotlib.pyplot as plt
import rl_utils


class PolicyNet(torch.nn.Module):
    def __init__(self, state_dim, hidden_dim, action_dim):
        super(PolicyNet, self).__init__()
        self.fc1 = torch.nn.Linear(state_dim, hidden_dim)
        self.fc2 = torch.nn.Linear(hidden_dim, action_dim)

    def forward(self, x):
        x = F.relu(self.fc1(x))
        return F.softmax(self.fc2(x), dim=1)


class ValueNet(torch.nn.Module):
    def __init__(self, state_dim, hidden_dim):
        super(ValueNet, self).__init__()
        self.fc1 = torch.nn.Linear(state_dim, hidden_dim)
        self.fc2 = torch.nn.Linear(hidden_dim, 1)

    def forward(self, x):
        x = F.relu(self.fc1(x))
        return self.fc2(x)


class PPO:
    ''' PPO算法,采用截断方式 '''
    def __init__(self, state_dim, hidden_dim, action_dim, actor_lr, critic_lr, lmbda, epochs, eps, gamma, device):
        self.actor = PolicyNet(state_dim, hidden_dim, action_dim).to(device)
        self.critic = ValueNet(state_dim, hidden_dim).to(device)
        self.actor_optimizer = torch.optim.Adam(self.actor.parameters(), lr=actor_lr)
        self.critic_optimizer = torch.optim.Adam(self.critic.parameters(), lr=critic_lr)
        self.gamma = gamma
        self.lmbda = lmbda
        self.epochs = epochs  # 一条序列的数据用来训练轮数
        self.eps = eps  # PPO中截断范围的参数
        self.device = device

    def take_action(self, state):  # 动作的选择没有发生改变和TRPO算法  策略梯度算法是一样的
        state = torch.tensor([state], dtype=torch.float).to(self.device)
        probs = self.actor(state)
        action_dist = torch.distributions.Categorical(probs)
        action = action_dist.sample()
        return action.item()

    def update(self, transition_dict):
        states = torch.tensor(transition_dict['states'], dtype=torch.float).to(self.device)
        actions = torch.tensor(transition_dict['actions']).view(-1, 1).to(self.device)
        rewards = torch.tensor(transition_dict['rewards'], dtype=torch.float).view(-1, 1).to(self.device)
        next_states = torch.tensor(transition_dict['next_states'], dtype=torch.float).to(self.device)
        dones = torch.tensor(transition_dict['dones'], dtype=torch.float).view(-1, 1).to(self.device)
        td_target = rewards + self.gamma * self.critic(next_states) * (1 - dones)
        td_delta = td_target - self.critic(states)
        advantage = rl_utils.compute_advantage(self.gamma, self.lmbda, td_delta.cpu()).to(self.device)
        old_log_probs = torch.log(self.actor(states).gather(1, actions)).detach()

        for _ in range(self.epochs):  # 策略更新  价值更新
            log_probs = torch.log(self.actor(states).gather(1, actions))
            ratio = torch.exp(log_probs - old_log_probs)
            surr1 = ratio * advantage  # 计算策略目标
            surr2 = torch.clamp(ratio, 1 - self.eps, 1 + self.eps) * advantage  # 截断
            actor_loss = torch.mean(-torch.min(surr1, surr2))  # PPO损失函数
            critic_loss = torch.mean(F.mse_loss(self.critic(states), td_target.detach()))
            self.actor_optimizer.zero_grad()
            self.critic_optimizer.zero_grad()
            actor_loss.backward()
            critic_loss.backward()
            self.actor_optimizer.step()
            self.critic_optimizer.step()


actor_lr = 1e-3
critic_lr = 1e-2
num_episodes = 500
hidden_dim = 128
gamma = 0.98
lmbda = 0.95
epochs = 10
eps = 0.2
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

env_name = 'CartPole-v0'
env = gym.make(env_name)
env.seed(0)
torch.manual_seed(0)
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.n
agent = PPO(state_dim, hidden_dim, action_dim, actor_lr, critic_lr, lmbda, epochs, eps, gamma, device)

return_list = rl_utils.train_on_policy_agent(env, agent, num_episodes)

episodes_list = list(range(len(return_list)))
plt.plot(episodes_list, return_list)
plt.xlabel('Episodes')
plt.ylabel('Returns')
plt.title('PPO on {}'.format(env_name))
plt.show()

mv_return = rl_utils.moving_average(return_list, 9)
plt.plot(episodes_list, mv_return)
plt.xlabel('Episodes')
plt.ylabel('Returns')
plt.title('PPO on {}'.format(env_name))
plt.show()

个人见解：（有不对的地方欢迎讨论）

PPO算法和TRPO算法最大的不同就是在寻找最优策略上，PPO算法使用截断法寻找最优策略，更方便简单，而TRPO算法是使用共轭梯度法和直线搜索方法，比较复杂

不管是PPO算法还是TRPO算法都是在Actor-Critic算法基础上的改进，TRPO算法是通过最优化方法找到一个信任区域，PPO算法是优化TRPO算法中寻找最优区域的方法，通过截断法来最优化目标。但是本质上还是用价值函数来指导策略函数以选取最有动作的策略。

Iteration 0:   0%|          | 0/50 [00:00<?, ?it/s]/usr/local/lib/python3.7/dist-packages/ipykernel_launcher.py:43: UserWarning: Creating a tensor from a list of numpy.ndarrays is extremely slow. Please consider converting the list to a single numpy.ndarray with numpy.array() before converting to a tensor. (Triggered internally at  ../torch/csrc/utils/tensor_new.cpp:201.)
Iteration 0: 100%|██████████| 50/50 [00:02<00:00, 19.41it/s, episode=50, return=183.200]
Iteration 1: 100%|██████████| 50/50 [00:03<00:00, 13.49it/s, episode=100, return=184.900]
Iteration 2: 100%|██████████| 50/50 [00:03<00:00, 12.64it/s, episode=150, return=200.000]
Iteration 3: 100%|██████████| 50/50 [00:03<00:00, 12.64it/s, episode=200, return=200.000]
Iteration 4: 100%|██████████| 50/50 [00:03<00:00, 12.81it/s, episode=250, return=200.000]
Iteration 5: 100%|██████████| 50/50 [00:03<00:00, 12.63it/s, episode=300, return=200.000]
Iteration 6: 100%|██████████| 50/50 [00:03<00:00, 12.83it/s, episode=350, return=200.000]
Iteration 7: 100%|██████████| 50/50 [00:03<00:00, 12.58it/s, episode=400, return=200.000]
Iteration 8: 100%|██████████| 50/50 [00:03<00:00, 12.78it/s, episode=450, return=200.000]
Iteration 9: 100%|██████████| 50/50 [00:03<00:00, 12.59it/s, episode=500, return=187.200]

在这里插入图片描述

12.4.2 倒立摆环境

倒立摆环境是一个连续的动作和车杆环境最大的区别就是动作的选取方面车杆环境使用的是伯努利分布（二项分布），而倒立摆是一个连续的动作，用的是高斯分布

import gym
import torch
import torch.nn.functional as F
import numpy as np
import matplotlib.pyplot as plt
import rl_utils


class PolicyNetContinuous(torch.nn.Module):
    def __init__(self, state_dim, hidden_dim, action_dim):
        super(PolicyNetContinuous, self).__init__()
        self.fc1 = torch.nn.Linear(state_dim, hidden_dim)
        self.fc_mu = torch.nn.Linear(hidden_dim, action_dim)
        self.fc_std = torch.nn.Linear(hidden_dim, action_dim)

    def forward(self, x):
        x = F.relu(self.fc1(x))
        mu = 2.0 * torch.tanh(self.fc_mu(x))
        std = F.softplus(self.fc_std(x))
        return mu, std


class ValueNet(torch.nn.Module):
    def __init__(self, state_dim, hidden_dim):
        super(ValueNet, self).__init__()
        self.fc1 = torch.nn.Linear(state_dim, hidden_dim)
        self.fc2 = torch.nn.Linear(hidden_dim, 1)

    def forward(self, x):
        x = F.relu(self.fc1(x))
        return self.fc2(x)

class PPOContinuous:
    ''' 处理连续动作的PPO算法 '''
    def __init__(self, state_dim, hidden_dim, action_dim, actor_lr, critic_lr,
                 lmbda, epochs, eps, gamma, device):
        self.actor = PolicyNetContinuous(state_dim, hidden_dim,
                                         action_dim).to(device)
        self.critic = ValueNet(state_dim, hidden_dim).to(device)
        self.actor_optimizer = torch.optim.Adam(self.actor.parameters(),
                                                lr=actor_lr)
        self.critic_optimizer = torch.optim.Adam(self.critic.parameters(),
                                                 lr=critic_lr)
        self.gamma = gamma
        self.lmbda = lmbda
        self.epochs = epochs
        self.eps = eps
        self.device = device

    def take_action(self, state):
        state = torch.tensor([state], dtype=torch.float).to(self.device)
        mu, sigma = self.actor(state)
        action_dist = torch.distributions.Normal(mu, sigma)
        action = action_dist.sample()
        return [action.item()]

    def update(self, transition_dict):
        states = torch.tensor(transition_dict['states'],
                              dtype=torch.float).to(self.device)
        actions = torch.tensor(transition_dict['actions'],
                               dtype=torch.float).view(-1, 1).to(self.device)
        rewards = torch.tensor(transition_dict['rewards'],
                               dtype=torch.float).view(-1, 1).to(self.device)
        next_states = torch.tensor(transition_dict['next_states'],
                                   dtype=torch.float).to(self.device)
        dones = torch.tensor(transition_dict['dones'],
                             dtype=torch.float).view(-1, 1).to(self.device)
        rewards = (rewards + 8.0) / 8.0  # 和TRPO一样,对奖励进行修改,方便训练
        td_target = rewards + self.gamma * self.critic(next_states) * (1 -
                                                                       dones)
        td_delta = td_target - self.critic(states)
        advantage = rl_utils.compute_advantage(self.gamma, self.lmbda,
                                               td_delta.cpu()).to(self.device)
        mu, std = self.actor(states)
        action_dists = torch.distributions.Normal(mu.detach(), std.detach())
        # 动作是正态分布
        old_log_probs = action_dists.log_prob(actions)

        for _ in range(self.epochs):
            mu, std = self.actor(states)
            action_dists = torch.distributions.Normal(mu, std)
            log_probs = action_dists.log_prob(actions)
            ratio = torch.exp(log_probs - old_log_probs)
            surr1 = ratio * advantage
            surr2 = torch.clamp(ratio, 1 - self.eps, 1 + self.eps) * advantage
            actor_loss = torch.mean(-torch.min(surr1, surr2))
            critic_loss = torch.mean(
                F.mse_loss(self.critic(states), td_target.detach()))
            self.actor_optimizer.zero_grad()
            self.critic_optimizer.zero_grad()
            actor_loss.backward()
            critic_loss.backward()
            self.actor_optimizer.step()
            self.critic_optimizer.step()

actor_lr = 1e-4
critic_lr = 5e-3
num_episodes = 2000
hidden_dim = 128
gamma = 0.9
lmbda = 0.9
epochs = 10
eps = 0.2
device = torch.device("cuda") if torch.cuda.is_available() else torch.device(
    "cpu")

env_name = 'Pendulum-v0'
env = gym.make(env_name)
env.seed(0)
torch.manual_seed(0)
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]  # 连续动作空间
agent = PPOContinuous(state_dim, hidden_dim, action_dim, actor_lr, critic_lr,
                      lmbda, epochs, eps, gamma, device)

return_list = rl_utils.train_on_policy_agent(env, agent, num_episodes)

episodes_list = list(range(len(return_list)))
plt.plot(episodes_list, return_list)
plt.xlabel('Episodes')
plt.ylabel('Returns')
plt.title('PPO on {}'.format(env_name))
plt.show()

mv_return = rl_utils.moving_average(return_list, 21)
plt.plot(episodes_list, mv_return)
plt.xlabel('Episodes')
plt.ylabel('Returns')
plt.title('PPO on {}'.format(env_name))
plt.show()

Iteration 0: 100%|██████████| 200/200 [00:22<00:00,  9.02it/s, episode=200, return=-1000.354]
Iteration 1: 100%|██████████| 200/200 [00:22<00:00,  8.78it/s, episode=400, return=-922.780]
Iteration 2: 100%|██████████| 200/200 [00:20<00:00,  9.63it/s, episode=600, return=-483.957]
Iteration 3: 100%|██████████| 200/200 [00:20<00:00,  9.80it/s, episode=800, return=-472.933]
Iteration 4: 100%|██████████| 200/200 [00:20<00:00,  9.54it/s, episode=1000, return=-327.589]
Iteration 5: 100%|██████████| 200/200 [00:20<00:00,  9.63it/s, episode=1200, return=-426.262]
Iteration 6: 100%|██████████| 200/200 [00:20<00:00,  9.73it/s, episode=1400, return=-224.806]
Iteration 7: 100%|██████████| 200/200 [00:21<00:00,  9.49it/s, episode=1600, return=-279.722]
Iteration 8: 100%|██████████| 200/200 [00:20<00:00,  9.62it/s, episode=1800, return=-428.538]
Iteration 9: 100%|██████████| 200/200 [00:20<00:00,  9.81it/s, episode=2000, return=-235.771]

在这里插入图片描述

PPO算法是TRPO算法的一种改进算法，它在实现上简化了TRPO中的复杂计算，并且它在实验中的性能大多数情况下会比TRPO更好，因此目前常被用作一种的基准算法。需要注意的是，TRPO和PPO算法都属于在线策略学习算法，即使优化目标中包含重要性采样的过程，但其只是用到了上一轮策略的数据，而不是过去所有策略的数据。

Python

torch.clamp()

函数原型：

torch.clamp(input, min=None, max=None, *, out=None)

clamp的是夹子，夹紧

所以这个函数的作用是将输入的input夹紧在min和max之间

当input在min和max之间的时候输出本身，小于min的时候输出min 大于max时，输出max

import torch
a = torch.randn(4)
b = torch.clamp(a, min=-0.5, max=0.5)
print(a, '\n', b)

输出为：

tensor([-1.7120,  0.1734, -0.0478, -0.0922])
tensor([-0.5000,  0.1734, -0.0478, -0.0922])

小点点~~

关注

20
点赞
踩
16

收藏

觉得还不错? 一键收藏
0
评论
PPO算法车杆环境倒立摆环境

上一章介绍的TRPO算法在很多场景上的应用都很成功，但是它的计算过程非常复杂，每一步更新的运算量非常大。于是，TRPO算法的改进版–PPO算法在2017年被提出，PPO基于TRPO的思想，但是其算法实现更加简单。并且大量试验结果表明，与TRPO算法相比，PPO能学习得一样好（甚至更快），这使得PPO成为非常流行的强化学习算法。如果我们想要尝试在一个新的环境中使用强化学习算法，那么PPO就属于可以首先尝试的算法。
复制链接

扫一扫