强化学习-PPO算法实现pendulum_强化学习 nan ppo

代码都是学习别人的，但我分享几点我踩过的大坑。

1.蒙特卡洛的V值

书上给的例子，是一次取一条轨迹，v=r+gamma*v 依次计算状态价值，这几乎是全部用蒙特卡洛方法的计算状态价值，并且没有对各条轨迹取均值，我想这种方法是极其不好的

2.样本不是独立同分布

由于1.中的原因，取到的样本不是独立同分布，把这种样本放入训练，可能会大幅影响训练效果。

（2022/12/29注 ：可见当时没学明白，ppo是on-policy 样本，1中也不是蒙特卡洛，还是TD-error）

3.代码写的太繁复。

俗话说的好，宁简勿繁，把太多方法封装成函数，在前期是不太好的行为，非常不便于调试，应当全部删去。

4.神经网络极易输出[nan]

可能是因为用了torch.Tensor()来转化向量，double型向量这使得他的内存占用高，改为torch.FloatTensor（）有明显改善。这一点极其重要，如果不用这个很可能根本没办法训练

                                                                训练效果

 代码如下

"""
"""
import torch.nn.functional as F
import torchvision.models as models
import retro
import hiddenlayer as hl
import torch
# import retro
import pandas as pd
import numpy as np
import gym
import torch.nn as nn
from torch.distributions import Normal
class DQBReplayer:
    def __init__(self,capacity):
        self.memory = pd.DataFrame(index=range(capacity),columns=['observation','action','reward','next_observation','done','step'])
        self.i=0
        self.count=0
        self.capacity=capacity
    def store(self,*args):
 
        self.memory.loc[self.i]=args
        self.i=(self.i+1)%self.capacity
        self.count=min(self.count+1,self.capacity)
    def sample(self,size=32):
        indics=np.random.choice(self.count,size=size)
 
        return (np.stack(self.memory.loc[indics,field]) for field in self.memory.columns)#为什么#是第indics行和feild列
    def clear(self):
        self.memory.drop(self.memory.index,inplace=True)
        self.count=0
        self.i=0
#
class PolicyNetwork(nn.Module):
    def __init__(self):
        super(PolicyNetwork, self).__init__()
        self.relu = nn.ReLU()
        self.fc1 = nn.Linear(3, 64)
        self.fc2 = nn.Linear(64, 256)
        self.fc_mu = nn.Linear(256, 1)
        self.fc_std = nn.Linear(256, 1)
        self.tanh = nn.Tanh()
        self.softplus = nn.Softplus()
 
 
 
    def forward(self, x):
        x = self.relu(self.fc1(x))
        x = self.relu(self.fc2(x))
        mu = 2 * self.tanh(self.fc_mu(x))
        std = self.softplus(self.fc_std(x)) + 1e-3
        return mu, std
 
    def select_action(self, state):
 
        with torch.no_grad():
            mu, std = self.forward(state)
            n = Normal(mu, std)
            action = n.sample()
        # print(" ac{:.1f},mu{},std{}".format( float(action),mu,std), end=" ")
        return np.clip(action.item(), -2., 2.)
 
 
class ValueNetwork(nn.Module):
    def __init__(self):
        super(ValueNetwork, self).__init__()
        self.relu = nn.ReLU()
        self.fc1 = nn.Linear(3, 64)
        self.fc2 = nn.Linear(64, 256)
        self.fc3 = nn.Linear(256, 1)
 
    def forward(self, x):
        x = self.relu(self.fc1(x))
        x = self.relu(self.fc2(x))
        x = self.fc3(x)
        return x
 
 
 
class PPO(nn.Module):
    def __init__(self):
        super(PPO,self).__init__()
        self.replayer=DQBReplayer(capacity=1000)
        self.gamma=0.99
        self.policy = PolicyNetwork().to(device)
        self.old_policy = PolicyNetwork().to(device)
        self.value = ValueNetwork().to(device)
        self.learn_step=0
        self.canvasl = hl.Canvas()
        self.history = hl.History()
 
 
 
if __name__ == "__main__":
    device=torch.device("cuda" if torch.cuda.is_available() else"cpu")
    env=gym.make("Pendulum-v0").unwrapped
 
    net=PPO().to(device)
    optim = torch.optim.Adam(net.policy.parameters(), lr=0.001)
    value_optim= torch.optim.Adam(net.value.parameters(), lr=0.001)
 
    for i in range(200000):
        state = env.reset()
        epoch_reward=0#每局游戏的累计奖励
        for step in range(200):
            # env.render()
            state_tensor = torch.FloatTensor(state).to(device)
            action=net.policy.select_action(state_tensor)
            next_state,r,done,info=env.step([action])
 
            reward = (r + 8.1) / 8.1
            epoch_reward+=reward
            net.replayer.store(state, action, reward, next_state, done,step)
            net.learn_step += 1
            state = next_state
 
        net.old_policy.load_state_dict(net.policy.state_dict())
        for K in range(10):
            sample_n = net.replayer.count
            states, actions, rewards, next_states, dones, steps = net.replayer.sample(32)
            states = torch.FloatTensor(states).to(device)
            next_states = torch.FloatTensor(next_states).to(device)
            actions = torch.FloatTensor(actions).unsqueeze(1).to(device)
            rewards = torch.FloatTensor(rewards).unsqueeze(1).to(device)
            with torch.no_grad():  # 为什么
                old_mu, old_std = net.old_policy(states)
                old_n = Normal(old_mu, old_std)
 
                value_target = rewards + net.gamma * net.value(next_states)
                advantage = value_target - net.value(states)
 
            mu, std = net.policy(states)
            n = Normal(mu, std)
            log_prob = n.log_prob(actions)
            old_log_prob = old_n.log_prob(actions)
            ratio = torch.exp(log_prob - old_log_prob)
            L1 = ratio * advantage
            L2 = torch.clamp(ratio, 0.8, 1.2) * advantage
            loss = torch.min(L1, L2)
            loss = - loss.mean()
            # writer.add_scalar('action loss', loss.item(), steps)
 
            optim.zero_grad()
            loss.backward()
            optim.step()
#clear
            value_loss = F.mse_loss(value_target, net.value(states))
            value_optim.zero_grad()
            value_loss.backward()
            value_optim.step()
        net.replayer.clear()
                # writer.add_scalar('value loss', value_loss.item(), steps)
 
        if i % 10 == 0 and i!=0:
            print('Epoch:{}, episode reward is {}'.format(i, epoch_reward))
            torch.save(net.policy.state_dict(), "pendulun_para\\reward"+str(epoch_reward//10)+'ppo-policy.para')
            # net.history.log((i * 200), avg_reward=epoch_reward/10)
            # with net.canvasl:
            #     net.canvasl.draw_plot(net.history["avg_reward"])
            epoch_reward = 0