【强化学习PPO算法】

  最近再改一个代码，需要改成PPO方式的，由于之前没有接触过此类算法，因此进行了简单学习，论文没有看的很详细，重点看了实现部分，这里只做简单记录。
  这里附上论文链接，需要的可以详细看一下。
   Proximal Policy Optimization Algorithms.

一、PPO算法

  PPO算法本质上是一个On-Policy的算法，它可以对采样到的样本进行多次利用，在一定程度上解决样本利用率低的问题，收到较好的效果。论文里有两种实现方式，一种是结合KL的penalty的，另一种是clip裁断的方法。大部分都是采用的后者，本文记录的也主要是后者的实现。

二、伪代码

  在网上找了一下伪代码，大概两类，前者是Open AI的，比较精炼，后者是Deepmind的，写的比较详细，在这里同时附上.

三、相关的简单理论

1.ratio

  这里的比例ratio，是两种策略下动作的概率比，而在程序实现中，用的是对动作分布取对数，而后使用e指数相减的方法，具体实现如下所示：

action_logprobs = dist.log_prob(action)
ratios = torch.exp(logprobs - old_logprobs.detach())

2.裁断

  其中，裁断对应的部分如下图所示：

  上述公式代表的含义如下：
  clip公式含义.

  这里我是这样理解的：
  (1)如果A>0,说明现阶段的(st,at)相对较好，那么我们希望该二元组出现的概率越高越好，即ratio中的分子越大越好，但是分母分子不能差太多，因此需要加一个上限;
  (2)如果A<0,说明现阶段的(st,at)相对较差，那么我们希望该二元组出现的概率越低越好，即ratio中的分子越小越好，但是分母分子不能差太多，因此需要加一个下限.

3.Advantage的计算

  论文里计算At的方式如下，在一些情况下可以令lamda为1；还有一种更常用的计算方式是VAE，这里不进行描述.。

  对应的代码块如下：

 def update(self, memory):
        # Monte Carlo estimate of rewards:
        rewards = []
        discounted_reward = 0
        for reward, is_terminal in zip(reversed(memory.rewards), reversed(memory.is_terminals)):
            if is_terminal:
                discounted_reward = 0
            discounted_reward = reward + (self.gamma * discounted_reward)
            rewards.insert(0, discounted_reward)

4.loss的计算

  这里的第一项，对应裁断项,需要计算ratio和Advantage，之后进行裁断；
  这里的第二项，对应的为对应的值的均方误差;
  这里的第三项，为交叉熵
  程序的实现如下所示：

surr1 = ratios * advantages
surr2 = torch.clamp(ratios, 1 - self.eps_clip, 1 + self.eps_clip) * advantages
loss = -torch.min(surr1, surr2) + 0.5 * self.MseLoss(state_values, rewards) - 0.01 * dist_entropy

四、算法实现

  这里算法的实现参考了一位博主
  PPO代码.

#!/usr/bin/python3
# -*-coding:utf-8 -*-

# @Time    : 2022/6/18 15:53
# @Author  : Wang xiangyu
# @File    : PPO.py
import torch
import torch.nn as nn
from torch.distributions import MultivariateNormal
import gym
import numpy as np

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")


class Memory:
    def __init__(self):
        self.actions = []
        self.states = []
        self.logprobs = []
        self.rewards = []
        self.is_terminals = []

    def clear_memory(self):
        # del语句作用在变量上，而不是数据对象上。删除的是变量，而不是数据。
        del self.actions[:]
        del self.states[:]
        del self.logprobs[:]
        del self.rewards[:]
        del self.is_terminals[:]


class ActorCritic(nn.Module):
    def __init__(self, state_dim, action_dim, action_std):
        super(ActorCritic, self).__init__()
        # action mean range -1 to 1
        self.actor = nn.Sequential(
            nn.Linear(state_dim, 64),
            nn.Tanh(),
            nn.Linear(64, 32),
            nn.Tanh(),
            nn.Linear(32, action_dim),
            nn.Tanh()
        )
        # critic
        self.critic = nn.Sequential(
            nn.Linear(state_dim, 64),
            nn.Tanh(),
            nn.Linear(64, 32),
            nn.Tanh(),
            nn.Linear(32, 1)
        )
        # 方差
        self.action_var = torch.full((action_dim,), action_std * action_std).to(device)

    def forward(self):
        # 手动设置异常
        raise NotImplementedError

    def act(self, state, memory):
        action_mean = self.actor(state)
        cov_mat = torch.diag(self.action_var).to(device)

        dist = MultivariateNormal(action_mean, cov_mat)
        action = dist.sample()
        action_logprob = dist.log_prob(action)

        memory.states.append(state)
        memory.actions.append(action)
        memory.logprobs.append(action_logprob)

        return action.detach()

    def evaluate(self, state, action):
        action_mean = self.actor(state)

        action_var = self.action_var.expand_as(action_mean)
        # torch.diag_embed(input, offset=0, dim1=-2, dim2=-1) → Tensor
        # Creates a tensor whose diagonals of certain 2D planes (specified by dim1 and dim2) are filled by input
        cov_mat = torch.diag_embed(action_var).to(device)
        # 生成一个多元高斯分布矩阵
        dist = MultivariateNormal(action_mean, cov_mat)
        # 我们的目的是要用这个随机的去逼近真正的选择动作action的高斯分布
        action_logprobs = dist.log_prob(action)
        # log_prob 是action在前面那个正太分布的概率的log ，我们相信action是对的 ，
        # 那么我们要求的正态分布曲线中点应该在action这里，所以最大化正太分布的概率的log， 改变mu,sigma得出一条中心点更加在a的正太分布。
        dist_entropy = dist.entropy()
        state_value = self.critic(state)

        return action_logprobs, torch.squeeze(state_value), dist_entropy


class PPO:
    def __init__(self, state_dim, action_dim, action_std, lr, betas, gamma, K_epochs, eps_clip):
        self.lr = lr
        self.betas = betas
        self.gamma = gamma
        self.eps_clip = eps_clip
        self.K_epochs = K_epochs

        self.policy = ActorCritic(state_dim, action_dim, action_std).to(device)
        self.optimizer = torch.optim.Adam(self.policy.parameters(), lr=lr, betas=betas)

        self.policy_old = ActorCritic(state_dim, action_dim, action_std).to(device)
        self.policy_old.load_state_dict(self.policy.state_dict())

        self.MseLoss = nn.MSELoss()

    def select_action(self, state, memory):
        state = torch.FloatTensor(state.reshape(1, -1)).to(device)
        return self.policy_old.act(state, memory).cpu().data.numpy().flatten()

    def update(self, memory):
        # Monte Carlo estimate of rewards:
        rewards = []
        discounted_reward = 0
        for reward, is_terminal in zip(reversed(memory.rewards), reversed(memory.is_terminals)):
            if is_terminal:
                discounted_reward = 0
            discounted_reward = reward + (self.gamma * discounted_reward)
            rewards.insert(0, discounted_reward)

        # Normalizing the rewards:
        rewards = torch.tensor(rewards, dtype=torch.float32).to(device)
        rewards = (rewards - rewards.mean()) / (rewards.std() + 1e-5)

        # convert list to tensor
        # 使用stack可以保留两个信息：[1. 序列] 和 [2. 张量矩阵] 信息，属于【扩张再拼接】的函数；
        old_states = torch.squeeze(torch.stack(memory.states).to(device), 1).detach()
        old_actions = torch.squeeze(torch.stack(memory.actions).to(device), 1).detach()
        old_logprobs = torch.squeeze(torch.stack(memory.logprobs), 1).to(device).detach()
		#这里即可以对样本进行多次利用，提高利用率
        # Optimize policy for K epochs:
        for _ in range(self.K_epochs):
            # Evaluating old actions and values :
            logprobs, state_values, dist_entropy = self.policy.evaluate(old_states, old_actions)

            # Finding the ratio (pi_theta / pi_theta__old):
            ratios = torch.exp(logprobs - old_logprobs.detach())

            # Finding Surrogate Loss:
            advantages = rewards - state_values.detach()
            surr1 = ratios * advantages
            surr2 = torch.clamp(ratios, 1 - self.eps_clip, 1 + self.eps_clip) * advantages
            loss = -torch.min(surr1, surr2) + 0.5 * self.MseLoss(state_values, rewards) - 0.01 * dist_entropy

            # take gradient step
            self.optimizer.zero_grad()
            loss.mean().backward()
            self.optimizer.step()

        # Copy new weights into old policy:
        self.policy_old.load_state_dict(self.policy.state_dict())


def main():
    ############## Hyperparameters ##############
    env_name = "Pendulum-v1"
    render = False
    solved_reward = 300  # stop training if avg_reward > solved_reward
    log_interval = 20  # print avg reward in the interval
    max_episodes = 10000  # max training episodes
    max_timesteps = 1500  # max timesteps in one episode

    update_timestep = 4000  # update policy every n timesteps
    action_std = 0.5  # constant std for action distribution (Multivariate Normal)
    K_epochs = 80  # update policy for K epochs
    eps_clip = 0.2  # clip parameter for PPO
    gamma = 0.99  # discount factor

    lr = 0.0003  # parameters for Adam optimizer
    betas = (0.9, 0.999)

    #############################################

    # creating environment
    env = gym.make(env_name)
    state_dim = env.observation_space.shape[0]
    action_dim = env.action_space.shape[0]

    memory = Memory()
    ppo = PPO(state_dim, action_dim, action_std, lr, betas, gamma, K_epochs, eps_clip)
    print(lr, betas)

    # logging variables
    running_reward = 0
    avg_length = 0
    time_step = 0

    # training loop
    for i_episode in range(1, max_episodes + 1):
        state = env.reset()
        for t in range(max_timesteps):
            time_step += 1
            # Running policy_old:
            action = ppo.select_action(state, memory)
            state, reward, done, _ = env.step(action)

            # Saving reward and is_terminals:
            memory.rewards.append(reward)
            memory.is_terminals.append(done)

            # update if its time
            if time_step % update_timestep == 0:
                ppo.update(memory)
                memory.clear_memory()
                time_step = 0
            running_reward += reward
            if render:
                env.render()
            if done:
                break

        avg_length += t+1

        # stop training if avg_reward > solved_reward
        if running_reward > (log_interval * solved_reward):
            print("########## Solved! ##########")
            torch.save(ppo.policy.state_dict(), './PPO_continuous_solved_{}.pth'.format(env_name))
            break

        # save every 500 episodes
        if i_episode % 500 == 0:
            torch.save(ppo.policy.state_dict(), './PPO_continuous_{}.pth'.format(env_name))

        # logging
        if i_episode % log_interval == 0:
            avg_length = int(avg_length / log_interval)
            running_reward = int((running_reward / log_interval))

            print('Episode {} \t Avg length: {} \t Avg reward: {}'.format(i_episode, avg_length, running_reward))
            running_reward = 0
            avg_length = 0


if __name__ == '__main__':
    main()

五、效果

  可以看到经过一段时间的训练，奖励有了一定升高.

六、感悟

  感悟是对改的项目的总结，和本文没有什么关系。
  这次改的项目参考了PPO的代码，架子基本也是搭好的，所以改起来也没有想象的那么困难。但应该是我第一次改代码，之前只是看代码，从来没有尝试改过那么多，可以感觉到看代码和改代码这两个能力间差的真的很多，写代码就更困难了emm，可以说经过这一次，可以更好的看到和别人的差距，不过对自己也有很大提高。在以后的学习中，还是需要多看多写，逐步提高。